JP2006228152A - Data analyzer - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a data analyzer, giving a solution having the continuity to the state of an initial pattern about a network constituted by data element groups as an object for analysis to be utilized for various purposes. <P>SOLUTION: In this data analyzer, the plurality of data elements are respectively associated with a data value of any of N stages (N is integers from 2 or more), and stored, the relational weighting information between the data elements is stored, one of the data elements is selected at random as an attentional data, and an input stimulation value related to the attentional data is computed on the basis of the relational weighting information between the attentional data and the other data element and the data value of the other data element. On the basis of the computed input stimulation value, a threshold determined depending on the current data value related to the attentional data and varied with every data value is compared with the input stimulation value to determine whether or not the data value is vaired, and if the data value is determined to be varied, the concerned data value is updated. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a data analysis apparatus using a neural network.

  As a data analysis method using a neural network, several methods have been conventionally known. Among them, the one proposed by Hopfield (JJHopfield) is that the coupling strength between the nodes is defined for a plurality of nodes, and the pattern including this node is a predetermined pattern group. It is to determine which is closest.

Here, the coupling strength between the nodes is set to be a so-called symmetric coupling in which the coupling strength from the node i to the node j is equal to the coupling strength from the node j to the node i. It is known that the result of activity propagation in a network of such nodes (symmetric coupling network) settles to a specific stable state (a fixed point that is an attractor of a dynamic system defined by activity propagation). In other words, Hopfield's proposed method determines which network's initial pattern is closest to which fixed point pattern (predetermined pattern) depending on which stable state (fixed point) the network eventually settles into. The calculation is performed (Non-Patent Document 1).
Hopfield, JJ (1982) Proc. Natl. Acad. Sci. USA, 79, 2554 Koulakov, AA, Raghavachari, S., Kepecs, A. & Lisman, JE (2002), Nat. Neurosci. 8, 775 Goldman MS, Levine JH, MajorG., Tank DW, Seung HS, "Robust persistent neural activity in a model integrator with multiple hesteretic dendrites per nuron", Cereb. Cortex 13, 1185-1195 (2003)

  However, in the conventional hop field method, the network analysis result is one of predetermined fixed point patterns. These fixed point patterns exist discretely in the solution space, and there is a problem that a solution that continuously changes with respect to the change of the initial pattern cannot be obtained.

  Therefore, an attempt has been made to realize a dynamic system having a continuous attractor by configuring a network with hysteresis units (units having hysteresis characteristics in change in value) (Non-patent Document 2). However, in this Kurarov method, the continuity of the pattern is not realized in the end, and the continuity regarding the scalar quantity (called activity) defined from the whole network is only realized.

  That is, no method has been found for providing a solution having continuity with respect to the initial pattern state of a network configured using a group of data elements to be analyzed. For this reason, any of the above conventional methods can only perform data analysis for limited applications.

  The present invention has been made in view of the above circumstances, and provides a solution having continuity with respect to the initial pattern state of a network configured using a data element group to be analyzed, and is used for a wide range of applications. It is an object of the present invention to provide a data analysis apparatus that can perform the above.

  Another object of the present invention is to provide a data analysis apparatus with improved network convergence.

  The present invention for solving the problems of the above-described conventional example is a data analysis apparatus, and stores a data value in any of N stages (N is an integer of 2 or more) in association with each of a plurality of data elements. And storage means for storing relation weighting information between the data elements, and selecting one of the data elements as attention data based on a predetermined rule, and weighting the relation between the attention data and another data element Means for calculating an input stimulus value related to the attention data based on information and data values of other data elements, and means for updating the data value of the attention data based on the calculated input stimulus value And determining whether or not to change the data value by comparing the threshold value which is determined according to the current data value related to the data of interest and is different for each data value and the input stimulus value. Means for updating the data value when it is determined to be changed, and means for scaling the data value so that the sum of the data values for each data element becomes a predetermined value. The data value for at least one of the data elements after being repeatedly subjected to the calculation of the input stimulus value and the update of the data value until a condition is satisfied is subjected to a predetermined process. It is characterized by.

  Further, the present invention for solving the problems of the conventional example is a data analyzing apparatus, wherein each of a plurality of data elements is a data value that is a continuous value from Ymin to Ymax (Ymin <Ymax). Are stored in association with each other, and storage means for storing relation weighting information between the data elements, one of the data elements is selected as attention data based on a predetermined rule, and the attention data and other data elements are selected. Based on the relationship weighting information and the data value of the other data element, the means for calculating the input stimulus value related to the attention data, the data of the attention data based on the calculated input stimulus value A means for updating a value, which is determined according to a current data value related to the data of interest, and compares a threshold value different for each data value with the input stimulus value to obtain a data value Decide whether to change, and if it is decided to change, means for updating the data value, and scale the data value so that the sum of the data values for each data element becomes a predetermined value The data value for at least one of the data elements after repeatedly performing the calculation of the input stimulus value and the update of the data value until a predetermined condition is satisfied, It is characterized by being subjected to predetermined processing.

  Here, the means for updating the data value is a first monotonically increasing function in which the input stimulus value is Ymin from 0 to X1min, continuously monotonically increases from X1min to X1max, and becomes Ymax above X1max. Using f1 and a second monotonically increasing function f2 in which the input stimulus value continuously increases monotonically between X2min and X2max and becomes Ymax above X2max, the input stimulus value for the data of interest ranges from X1min to X2max. In the meantime, with respect to the data value Y of the attention data before the update, the data value is updated to f2 (I) when the input stimulus value I exceeds I2 with reference to I2 where Y = f2 (I2) Referring to the data value Y of the attention data before update, referring to I1 where Y = f1 (I1), and updating the data value to f1 (I) when the input stimulus value I is less than I1 It is good.

  According to another aspect of the present invention, there is provided storage means for storing a plurality of discrete values or continuous values of data values in association with each of the plurality of data elements, and storing relationship weighting information between the data elements. A data analysis method using a computer, wherein one of the data elements is selected as attention data based on a predetermined rule, relation weighting information between the attention data and another data element, and other data A step of calculating an input stimulus value related to the attention data based on the data value of the element; and a means for updating the data value of the attention data based on the calculated input stimulus value. A threshold value that is determined according to the current data value related to the data and is different for each data value is compared with the input stimulus value to determine whether to change the data value. A predetermined condition is satisfied when the data value is updated, and the step of scaling the data value so that the sum of the data values for each data element becomes a predetermined value. The data value for at least one of the data elements after the repeated execution is subjected to a predetermined process.

  Furthermore, a program according to another aspect of the present invention stores a plurality of discrete values or continuous data values in association with each of a plurality of data elements, and stores relationship weighting information between the data elements. A computer having means for selecting one of the data elements as attention data based on a predetermined rule, relation weighting information between the attention data and another data element, a data value of the other data element, And a means for updating the data value of the attention data based on the calculated input stimulus value based on the calculated input stimulus value, and a current value related to the attention data. A threshold value that differs for each data value and the input stimulus value is compared to determine whether to change the data value. If the predetermined value is satisfied, the procedure for updating the data value and the procedure for scaling the data value so that the sum of the data values for each data element becomes a predetermined value. It is characterized by being repeatedly executed.

  Embodiments of the present invention will be described with reference to the drawings. As shown in FIG. 1, the data analysis apparatus according to the embodiment of the present invention includes a control unit 11, a storage unit 12, a data input unit 13, and a result output unit 14.

  The control unit 11 can be realized by a CPU or the like, and executes data analysis processing according to a program stored in the storage unit 12. The specific contents of this process by the control unit 11 will be described in detail later. The storage unit 12 can be realized using a memory element or a disk device. The storage unit 12 stores a program executed by the control unit 11. The storage unit 12 also operates as a work memory that holds various data used in the data analysis processing of the control unit 11.

  The data input unit 13 and the result output unit 14 are input / output interfaces that acquire data from, for example, a data storage that stores data to be analyzed, and output the analysis results to the data storage or display.

  Next, data analysis processing by the control unit 11 of the present embodiment will be described. In the present embodiment, an interconnection network type data structure is defined, and a data value storage unit having multi-stage hysteresis input / output characteristics is associated with each node. In the present embodiment, the data to be analyzed includes a plurality of data elements. Specifically, if the data to be analyzed is a sentence, the data element can be defined as an individual word.

  The multistage hysteresis input / output characteristics of each node are as follows. First, as shown in FIG. 2A, in the simple hysteresis characteristic (single hysteresis characteristic), the data value Y is either Y = 0 or Ymax. A lower threshold value Qmin and an upper threshold value Qmax are determined. Assuming that the data value Y is initially initialized to “0”, the state of “0” is maintained until the input I exceeds the upper threshold value Qmax. When the input I exceeds the upper threshold value Qmax, the data value Y becomes Ymax. After that, while the input I exceeds the lower threshold value Qmin, the value of Ymax is maintained. When the input I falls below the lower threshold value Qmin, the data value Y becomes “0”.

  Multi-stage hysteresis is obtained by accumulating this single hysteresis characteristic in multiple stages. Specifically, in this multi-stage hysteresis characteristic, each node takes one of N-stage values from Y = 0 to Y = Ymax. Specifically, as shown in FIG. 2B, a lower threshold value Qimin and an upper threshold value Qimax are defined for each stage.

  When the data value is reset to “0” and the input exceeds the upper threshold value Q1max and does not exceed the upper threshold value Q2max in the second stage (when Q1max <I <Q2max), The output is the first stage output Y = Ymax / N. If the input I changes from here to below the lower threshold value Q1min of the first stage, the data value Y becomes “0”.

  When the data value is Y = Ymax / N and the value of the input I exceeds the upper threshold value Q2max in the second stage and does not exceed the upper threshold value Q3max in the third stage (Q2max <I <In the case of Q3max), the data value Y is the second stage value, Y = 2 × Ymax / N.

  As shown in FIG. 3, the control unit 11 generates an element database in which information (for example, a character string of a word) that identifies a data element and a data value are associated with each data element, and stores the element database in the storage unit 12. . Further, as illustrated in FIG. 4, the control unit 11 generates a relation weighting database in which a relation weight value is set for each pair of data elements, and stores the relation weighting database in the storage unit 12. In the example of FIG. 4, the lower triangular part is omitted because the lower triangular part is set to a value obtained by folding the upper triangular part with a diagonal line (that is, Tij = Tji).

  Further, in the present embodiment, the storage unit 12 stores the lower threshold value and the upper threshold value of each step for defining the multistep hysteresis characteristics.

  The control unit 11 performs a data analysis process based on the information stored in the storage unit 12. In this embodiment, such a hysteresis unit is made to correspond to each node in the interconnection network, and active propagation is performed by a “dynamic system” determined in the following steps. In the network comprising the data value storage units of the present embodiment, the attractor (fixed point state) obtained by the active propagation continuously depends on the initial state.

  Hereinafter, the process of activity propagation in the control unit 11 will be described with reference to FIG. One data element is selected as attention data based on a predetermined rule among a plurality (M) of data elements included in the data to be processed (S1). This rule (selection rule) is a rule for selecting at random, for example. Here, it is assumed that the i-th data element is selected. Then, the input stimulus value for updating the data value associated with the attention data is calculated by the following equation (1) (S2).

  Here, Tij is a value of the relation weight between the i-th data element and the j-th data element, and is read from the relation weight database and used. Oj is a data value stored in the element database in association with the jth data element.

  Next, the data value associated with the i-th data element that is the data of interest is determined (S3), and the element database is updated. Specifically, the determination of the data value of the present embodiment is performed as follows.

  That is, it is checked which level n of the multilevel values the data value associated with the data of interest is. Here, since Ymax / N = ΔY is changed every step, the value n of the step can be obtained by dividing the data value by ΔY (that is, since Y = n × ΔY, this is expressed as ΔY. To obtain n). Next, the lower threshold value θmin and the upper threshold value θmax corresponding to the obtained stage are acquired.

  Then, the input stimulus value I calculated in the process S2 is compared with the upper threshold value θmax. If I> θmax, the data value Y is updated to (n + 1) × ΔY, and the process S3 is recursively performed. Run it. In the uppermost stage Y = Ymax, the upper threshold value is set larger than the maximum value of the input stimulus value so that the data value does not exceed the maximum value. Further, the lower threshold value θmin is compared with the input stimulus value I. When I ≦ θmin, the data value Y is updated to (n−1) × ΔY, and the process S3 is recursively executed. Also in this case, in the lowest stage Y = Ymin, the lower threshold value is set smaller than the minimum value of the input stimulus value so that the data value does not fall below the minimum value.

  If θmin <I ≦ θmax, the process proceeds to step S4.

を演算し、データ値Ｙiに代えて、このデータ値Ｙ′iを記憶部１２に上書きして格納する。このスケーリングによって、ネットワークの収束性が向上される。 The control unit 11 scales the data values so that the sum of the data values for each data element becomes a predetermined value C (S4). That is, if the i-th data value is Yi,

And the data value Y′i is overwritten and stored in the storage unit 12 instead of the data value Yi. This scaling improves network convergence.

を生成する（Ｓ５）。そして記憶部１２に前回生成したベクトルが格納されているときには、当該前回生成したベクトルと当該ベクトルとの差を算出し、当該算出された差のベクトルがゼロベクトル（その要素がすべて「０」となっているベクトル）か否か、つまり、

となっているか否かを調べ（Ｓ６）、ゼロベクトルでない場合（Ｎｏの場合）は、処理Ｓ１に戻って処理を繰返す。ここではデータ値の更新後において、各データ要素に関係するデータ値が変化しなくなったか否かを調べ、変化している間は処理Ｓ１に戻るようにしているのである。 After updating and scaling the data values associated with the data of interest in this way, a vector defined by the set of data values for each data element

Is generated (S5). When the previously generated vector is stored in the storage unit 12, the difference between the previously generated vector and the vector is calculated, and the calculated difference vector is a zero vector (all elements are “0”). Or not), that is,

(S6), if it is not a zero vector (in the case of No), the process returns to the process S1 and the process is repeated. Here, after the data value is updated, it is checked whether or not the data value related to each data element is changed, and the process returns to the process S1 while the data value is changed.

  If the calculated difference vector is a zero vector in the processing S6 (Yes), the processing is terminated. At this time, a set of information and data values specifying the data elements stored in the element database is provided as a result of analysis to a later predetermined process. The state stored in the element database represents a state in which the input / output of each data element is balanced (so-called dynamic system attractor). This state corresponds to a state called “fixed point”.

  In addition, in the calculation in the process S3 described here, it is necessary to execute the calculation process recursively. That is, the data value may change by a plurality of stages. In this way, when changing over a plurality of stages, the control unit 11 repeatedly performs a comparison operation between the threshold value and the input stimulus value. If the number N of stages is increased, the calculation speed is reduced. Therefore, in this embodiment, instead of discretely setting the data value to N stages, the data value is handled as a continuous value assuming the limit when N is infinite. The continuous value here is expressed as a discrete value as the internal representation of the computer, but in theory it is called a continuous value in the sense that it does not take a stepwise value. It is.

In the case of a continuous value, as shown in FIG. 6, the input stimulus value is Ymin from 0 to X1min, continuously increases monotonically from X1min to X1max, and becomes Ymax when X1max or more. Data value update processing is performed using a monotonically increasing function f1 and a second monotonically increasing function f2 in which the input stimulus value continuously increases monotonically between X2min and X2max and becomes Ymax when X2max or more. That is,
(1) When the input stimulus value I for the data of interest exceeds X2max, that is, when X2max <I:
The data value is Y = Ymax.
(2) When the input stimulus value I for the data of interest is X1 min or less, that is, when X1min <I:
The data value is Y = Ymin.
(3) When the input stimulus value I for the data of interest is between X1min and X2max, that is, when X1min <I ≦ X2max:
Regarding the data value Y of the attention data before update, referring to I2 where Y = f2 (I2), if the input stimulus value I exceeds the I2, the data value is updated to f2 (I),
With respect to the data value Y of the attention data before update, with reference to I1 where Y = f1 (I1), when the input stimulus value I is less than I1, the data value is updated to f1 (I).

として定義できる。この場合、Ｙ＝ｆ２（Ｉ２）、及びＹ＝ｆ２（Ｉ１）を、それぞれ、Ｉ２，Ｉ１について解いた解が

となる。従って、上記データ値の更新処理は、次のようになる。すなわち、
（１）注目データに対する入力刺激値ＩがＸ２maxを超える場合、つまりＸ２max＜Ｉの場合：
データ値Ｙ＝Ｙmaxとする。
（２）注目データに対する入力刺激値ＩがＸ１min以下の場合、つまりＸ１min＜Ｉの場合：
データ値Ｙ＝Ｙminとする。
（３）注目データに対する入力刺激値ＩがＸ1minからＸ2maxまでの間の場合、つまりＸ１min＜Ｉ≦Ｘ２maxの場合：
更新前の注目データのデータ値Ｙについて、入力刺激値ＩがＹ／α＋Ｘ２minを超える場合に、データ値をｆ２（Ｉ）＝α（Ｉ−Ｘ２min）に更新する。また更新前の注目データのデータ値Ｙについて、入力刺激値ＩがＹ／α＋Ｘ１min未満となる場合に、データ値をｆ１（Ｉ）＝α（Ｉ−Ｘ１min）に更新する。この方法によると、判断処理の回数を制限でき、計算量を低減できる。 Here, the first and second monotonically increasing functions f1 and f2 are

Can be defined as In this case, Y = f2 (I2) and Y = f2 (I1) are solved for I2 and I1, respectively.

It becomes. Therefore, the data value update process is as follows. That is,
(1) When the input stimulus value I for the data of interest exceeds X2max, that is, when X2max <I:
The data value is Y = Ymax.
(2) When the input stimulus value I for the data of interest is X1 min or less, that is, when X1min <I:
The data value is Y = Ymin.
(3) When the input stimulus value I for the data of interest is between X1min and X2max, that is, when X1min <I ≦ X2max:
When the input stimulus value I exceeds Y / α + X2min for the data value Y of the attention data before update, the data value is updated to f2 (I) = α (I−X2min). When the input stimulus value I is less than Y / α + X1min for the data value Y of the attention data before update, the data value is updated to f1 (I) = α (I−X1min). According to this method, the number of determination processes can be limited, and the amount of calculation can be reduced.

  Next, the operation of the data analysis apparatus according to this embodiment will be described. Here, an example will be described in which a normal document search process for searching a document including a desired keyword from a document group and a similar document search process for searching a similar document similar to a key document from the document group are described.

  In performing these processes, first, the following process is performed on a document group to be searched. That is, the control unit 11 retrieves data of a document group stored in a storage (not shown) via the data input unit 13, and for each of the P documents D1, D2, ..., DP, Q search terms w1, Extract w2,..., wQ. Here, the search term is extracted by removing a predetermined general-purpose word (so-called stop word) from a word group included in each document, and performing a stemming process (a process for extracting a stem part excluding use and ending change). It is what I did. Note that the stemming process is a process in which words such as child, childhood, and children are treated as the same “child” in consideration of ending changes and plural word changes.

とする。ここでａｉｊは、文書Ｄｊにおける単語ｗｉの出現回数を表し、ｂｉは、文書群全体において単語ｗｉを含む文書の数を表す。またＰは、文書群内の全文書の数である。 Next, the importance dij of each index word wi in each document Dj is defined by the TFIDF method (Term Frequency Inverse Document Frequency). That is,

And Here, aij represents the number of appearances of the word wi in the document Dj, and bi represents the number of documents including the word wi in the entire document group. P is the number of all documents in the document group.

  Next, using the index word as a data element, a storage area (element database) for storing information for specifying the index word and a data value is secured in the storage unit 12. Further, an area (relation weight database) for storing relation weights between index words is secured in the storage unit 12. Here, the relational weight is determined based on a covariant learning rule. That is, associating words that co-occur in the same document, the connection weight Tij between the words wi and wj is defined as follows.

  Here, when the normal document search process is performed, the TFIDF method is used based on the number of occurrences of the index word w included in the sentence and the appearance frequency of the index word in the document group for the sentence (query sentence) serving as a search key. Defines a vector q of the importance of the index word wi in the query sentence.

  Then, the importance value of each index word wi in the query sentence is stored in the element database of the storage unit 12 as the initial value of the data value associated with the corresponding index word. The control unit 11 performs the activity propagation process shown in FIG. When the data value stored in the element database is no longer updated, the control unit 11 reads out this data value set as an output (output vector). Note that the vector dj is a vector value in which the importance of each word is arranged for the j-th document, and the vector q is an output vector after the activity propagation process.

を演算する。そして、文書ごとに演算されたコサイン値（ｊ番目の文書についてＣ（ｊ））の大きい文書を降順に配列して検索結果として結果出力部１４を介して利用者に提示する。 The control unit 11 then calculates a cosine value related to the inner product of the output vector and the index word importance set for each document.

Is calculated. Then, documents having a large cosine value (C (j) for the j-th document) calculated for each document are arranged in descending order and presented as search results to the user via the result output unit 14.

  When similar document search processing is performed, instead of the vector q of the importance of the index word wi in the query sentence, the vector q of the importance of the index word wi in the key document is used, and processing similar to the above is performed. Should be done.

  In this way, by pre-processing the search key, such as a keyword or key document, by the processing shown in FIG. 5, for example, a document that does not appear in the search key but contains a related word actually It will be searched.

  In the data analysis apparatus according to the present embodiment, in order to use such activity propagation, a solution having continuity with respect to the state of the initial pattern of the network configured using the data element group to be analyzed is obtained. give. For this reason, it can be used for a wider range of applications than a network that converges into a solution that can be determined in a discontinuous manner.

  Furthermore, in the present embodiment, in the activity propagation process, data value scaling is performed (normalization is performed by dividing each data value by the sum value and multiplying it by C to obtain a constant sum C). Therefore, the convergence is improved without damaging the network structure. In this case, if the convergence is higher when the scaling is not performed based on the initial value of the data value, the user may be able to appropriately select whether or not to perform the scaling. When the user selects not to perform scaling, the control unit 11 proceeds to the process S5 after the process S3 without executing the process S4 in the process of FIG.

1 is a configuration block diagram of a data analysis apparatus according to an embodiment of the present invention. It is explanatory drawing showing the example of the update method of the data value for every data element. It is explanatory drawing showing the example of the content of an element database. It is explanatory drawing showing the example of the content of a relation weight database. It is a flowchart figure showing the processing example of activity propagation. It is explanatory drawing showing the example of the update method of a data value.

Explanation of symbols

  11 control unit, 12 storage unit, 13 data input unit, 14 result output unit.

Claims (4)

Storage means for associating and storing data values that are continuous values from Ymin to Ymax (Ymin <Ymax) for each of the plurality of data elements, and storing relationship weighting information between the data elements;
One of the data elements is selected as attention data based on a predetermined rule, and the attention data is determined based on relation weighting information between the attention data and another data element and the data value of the other data element. Means for calculating an input stimulus value related to
Based on the calculated input stimulus value, a means for updating the data value of the data of interest, which is determined according to the current data value related to the data of interest, and a different threshold for each data value; A means for comparing the input stimulus value to determine whether or not to change the data value, and, when determined to change, means for updating the data value;
Means for scaling the data values such that the sum of the data values for each data element is a predetermined value;
Including
The data value for at least one of the data elements after repeatedly executing the calculation of the input stimulus value and the update of the data value until a predetermined condition is satisfied is provided for a predetermined process. A data analyzer characterized by that. The data analysis device according to claim 1,
The means for updating the data value is:
When the input stimulus value is 0 to X1min, it is Ymin, and continuously increases monotonically from X1min to X1max. When the input stimulus value is X1max or more, the first monotonically increasing function f1 becomes Ymax, and the input stimulus value is from X2min to X2max. And a second monotonically increasing function f2 that continuously increases monotonically during the period, and becomes Ymax above X2max,
When the input stimulus value for the data of interest is between X1min and X2max,
Regarding the data value Y of the attention data before update, referring to I2 where Y = f2 (I2), if the input stimulus value I exceeds the I2, the data value is updated to f2 (I),
Regarding the data value Y of the attention data before update, referring to I1 where Y = f1 (I1), when the input stimulus value I is less than this I1, the data value is updated to f1 (I). Characteristic data analysis device. Using a computer comprising storage means for storing continuous weight data values in association with each of a plurality of data elements, and storing relationship weighting information between the data elements,
One of the data elements is selected as attention data based on a predetermined rule, and the attention data is determined based on relation weighting information between the attention data and another data element and the data value of the other data element. Calculating an input stimulus value related to
Based on the calculated input stimulus value, a means for updating the data value of the data of interest, which is determined according to the current data value related to the data of interest, and a different threshold for each data value; Comparing the input stimulus value to determine whether to change the data value, and if it is determined to change, updating the data value; and
Scaling the data values such that the sum of the data values for each data element is a predetermined value;
Is repeatedly executed until a predetermined condition is satisfied, and the data value for at least one of the data elements after the repeated execution is subjected to a predetermined process. A computer having storage means for storing a plurality of discrete values or continuous values of data values in association with each of a plurality of data elements and storing relation weighting information between the data elements,
One of the data elements is selected as attention data based on a predetermined rule, and the attention data is determined based on relation weighting information between the attention data and another data element and the data value of the other data element. A procedure for calculating an input stimulus value related to
Based on the calculated input stimulus value, a means for updating the data value of the data of interest, which is determined according to the current data value related to the data of interest, and a different threshold for each data value; Comparing the input stimulus value to determine whether to change the data value, and if it is determined to change, a procedure to update the data value;
A procedure for scaling the data values such that the sum of the data values for each data element is a predetermined value;
Is repeatedly executed until a predetermined condition is satisfied.

Images