JP5011529B2 - Data processing apparatus, data processing method, and program - Google Patents

Description

  The present invention relates to a data processing device, a data processing method, and a program, and more particularly, to a data processing device, a data processing method, and a program for constructing an intended vector set according to learning vectors.

  Patent Document 1 below discloses a self-organizing relationship (SOR) network. This self-organizing relationship network includes a competitive layer in which a predetermined number of neuron models are arranged. When a learning vector having a positive evaluation is input, the reference vectors of some or all of the neuron models are When a learning vector that approaches the learning vector and has a negative evaluation is inputted, the reference vectors of some or all of the neuron models are separated from the learning vector. According to the technology, for example, a learning vector composed of a pair of input data and output data is sequentially input to a self-organization relation network together with evaluation of the pair, thereby constructing a set of desirable input / output data pairs as a reference vector group. It becomes possible.

  However, according to the self-organization relationship network, the number of neuron models arranged in the competitive layer, that is, the number of reference vectors is fixed, and depending on the set of input learning vectors, a desired data set may be used as a reference vector group. It can be difficult to build properly.

In this regard, according to the ESOM (Evolving Self-Organizing Map) described in Non-Patent Document 1 below, the number of reference vectors can be increased by an input learning vector. That is, according to the ESOM, when a learning vector is input, if a reference vector close to the learning vector already exists, the reference vector and the reference vector in the vicinity thereof are updated according to the learning vector as in the prior art. If no reference vector exists, a reference vector that matches the learning vector is newly added to the ESOM. In this way, the number of reference vectors is increased as necessary, and the network is changed so that a desirable data set can be suitably constructed.
JP 2000-122991 A “On-line pattern analysis by evolving self-organizing maps”, Da Deng and Nikola Kasabov, Neurocomputing, Vol. 51, 87-103, 2003

  However, Non-Patent Document 1 does not specifically disclose a method for reducing the number of reference vectors, and therefore learning is performed when the relationship between input data and output data included in a learning vector changes over time. There is a problem that even if the property of the vector changes, the reference vector updated based on the past learning vector remains as it is. Therefore, according to ESOM, there is a problem that the reference vector group does not always coincide with a desired data set based on the current learning vector.

  The present invention has been made in view of the above background, and the object thereof is to suitably re-use an intended vector set as a reference vector group even when the properties of sequentially input learning vectors change over time. An object of the present invention is to provide a data processing apparatus, a data processing method, and a program that can be constructed.

  In order to solve the above problems, a data processing apparatus according to the present invention includes a reference vector storage unit that stores one or more reference vectors, a learning vector that is sequentially input, and each of the references that is stored in the reference vector storage unit A distance calculating means for calculating a distance to the vector; and a reference vector determined based on the learning vector when the minimum distance calculated by the distance calculating means is larger than a predetermined value in the reference vector storage means. Reference vector adding means for additionally storing, and at least a part of the one or more reference vectors stored in the reference vector storage means when the minimum value of the distance calculated by the distance calculating means is less than or equal to a predetermined value Reference vector updating means for making the learning vector approach or separate according to the evaluation of the learning vector, The activity level is stored in association with each reference vector stored in the reference vector storage means, and the activity level is updated based on the distance between the reference vector and the learning vector and the evaluation of the learning vector. And storing and updating means, and reference vector deleting means for deleting each reference vector from the reference vector storage means in accordance with the degree of activity associated with each reference vector.

  The data processing method according to the present invention includes a distance calculating step of calculating a distance between a learning vector that is sequentially input and each of the reference vectors stored in a reference vector storage unit that stores one or more reference vectors. A reference vector adding step for adding and storing a reference vector determined based on the learning vector in the reference vector storage means when a minimum value of the distance calculated in the distance calculating step is larger than a predetermined value; When the minimum value of the distance calculated in the distance calculating step is less than or equal to a predetermined value, at least a part of the one or more reference vectors stored in the reference vector storage unit is set to the learning vector. A reference vector update step for approaching or separating the reference vector according to the evaluation of the reference vector, and stored in the reference vector storage means Based on the distance between each reference vector and the learning vector and the evaluation with respect to the learning vector, an activity update step for updating the activity of the reference vector, and depending on the activity of each reference vector, each reference And a reference vector deleting step of deleting the vector from the reference vector storage means.

  Further, the program according to the present invention includes a reference vector storage unit that stores one or more reference vectors, a distance calculation that calculates a distance between a sequentially input learning vector and each of the reference vectors stored in the reference vector storage unit. Means for adding a reference vector determined on the basis of the learning vector to the reference vector storage means when the minimum distance calculated by the distance calculation means is larger than a predetermined value; When the minimum distance calculated by the distance calculation means is less than or equal to a predetermined value, at least a part of the one or more reference vectors stored in the reference vector storage means is set to the learning vector with respect to the learning vector. Reference vector update means for making them approach or separate in accordance with the evaluation for, stored in the reference vector storage means An activity level is stored in association with each reference vector, and an activity level update unit for updating the activity level based on a distance between the reference vector and the learning vector and an evaluation of the learning vector, and each reference A program for causing a computer to function as reference vector deleting means for deleting each reference vector from the reference vector storage means in accordance with the degree of activity associated with a vector. This program may be stored in a computer-readable information storage medium such as a CD-ROM or a DVD-ROM.

  According to the present invention, the activity is stored in association with each reference vector. The degree of activity is updated based on the distance from the learning vector that is sequentially input and the evaluation of the learning vector. For example, the activity storage update unit may change the activity greatly as the distance between the reference vector and the learning vector is shorter. The activity storage update means may increase the activity when the evaluation with respect to the learning vector is positive, and decrease the activity when the evaluation is negative. In the present invention, since the reference vector is deleted from the reference vector storage means according to the activity stored / updated in this way, a reference vector that is not based on the property of the currently input learning vector is stored in the reference vector storage means. It is possible to prevent the problem of being continued. Therefore, even if the properties of the learning vectors input sequentially change with time, the intended vector set can be suitably reconstructed as a reference vector group.

  The reference vector adding means refers to the reference vector when the minimum distance calculated by the distance calculating means is larger than the predetermined value and the evaluation with respect to the learning vector is not less than a predetermined value. You may make it add and memorize | store in a vector memory | storage means.

  Further, the coupling strength of the pair is stored in association with each pair of the reference vectors stored in the reference vector storage means, and the coupling strength is updated according to the distance between each of the pair and the learning vector. Further, it is possible to further include a bond strength memory updating means. At this time, in addition to the distance between each of the pairs and the learning vector, the connection strength storage updating unit may further update the connection strength according to the evaluation with respect to the learning vector.

  In this case, the reference vector update means makes the reference vector related to the minimum value and a reference vector in the vicinity of the reference vector determined based on the coupling strength close to or separated from the learning vector. You may let me. The reference vector deleting unit may also delete the joint strength associated with the reference vector when deleting the reference vector from the reference vector storage unit.

  Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

  The data processing apparatus according to the present embodiment is configured around a known computer system, and conceptually implements an improved self-organization relation network on a computer. That is, in this data processing apparatus, in a learning mode, a large number of learning vectors are sequentially input to a neural network having a competitive layer in which an arbitrary number of one or more units (neuron models) are arranged. The learning process for reflecting the contents of the above in the reference vector of each unit is executed. At this time, in this embodiment, the learning vector and the reference vector have an input data component and an output data component, and the learning vector is input to the neural network together with the evaluation value of the input data and output data pair. Is done. When a learning vector for positive evaluation, that is, a pair of input data and output data having a positive evaluation value, is input, reference vectors of some or all of the units approach the learning vector. Conversely, when a learning vector for negative evaluation, that is, a pair of input data and output data having a negative evaluation value, is input, the reference vectors of some or all units are separated from the learning vector.

  In the execution mode, only the input data is input to the neural network, and the similarity (for example, a decreasing function of the distance (Euclidean distance)) between the input data and the input data component of each reference vector is calculated. The output data is calculated by weighted addition of the components of the output data of each reference vector using this similarity.

  In an example applied to a control system, input data is data of a control error (a difference between a controlled variable and its target value) related to a controlled object, output data is data of a manipulated variable related to an operated object, and an evaluation value is In the case of the control error, the value indicates the good control when the operation amount is applied to the operation target (for example, a value indicating whether or not the control amount has converged). In this way, in the learning mode, a control error and manipulated variable pair group for realizing preferable control is acquired as the reference vector group. In the execution mode, by inputting data of an arbitrary control error, operation amount data for realizing preferable control is output. Thus, according to the present embodiment, a preferable input / output relationship of the target system can be obtained by trial and error, and then, by providing input data, it is possible to output preferable output data corresponding thereto.

  This data processing apparatus performs information processing based on the self-organization relation network as described above, and has the following features. That is, this data processing apparatus has a function of adding or deleting a unit on the competitive layer when the preferable input / output relationship changes and the nature of the learning data changes accordingly. Thereby, even if the preferable input / output relationship changes, it can be appropriately acquired as a reference vector group.

  Further, the strength of coupling between units, that is, the coupling strength is managed, and in the learning mode, the units coupled with a strength larger than a predetermined value (for example, zero) are collectively updated. In the execution mode, for each subset of units that are connected directly or indirectly with a strength of a predetermined value or more, the similarity between the input data and the component of the input data of each reference vector belonging to the subset is calculated. The output data is calculated by weighted addition of the components of the output data of each reference vector using this similarity. Thereby, even when there are a plurality of preferable output data for the input data, they can be output.

1 to 3 conceptually show the processing of the data processing apparatus. In these figures, a reference vector space having an axis of input data u and an axis of output data v is shown, a white circle indicates a reference vector, and a broken line indicates a link having a coupling strength value greater than zero. It shows that it exists in between. Black circles indicate learning vectors. The learning vector x is a two-dimensional vector having input data p and output data q as components as shown in the following equation (1), and the reference vector w _i is input data as shown in the following equation (2). It is a two-dimensional vector having u _i and output data v _i as components. Here, it is assumed that N units are arranged in the competitive layer as an initial state, and N reference vectors w _i initially exist.

As shown in FIG. 1, when a learning vector xa is input at a position within a predetermined distance (receptive field radius) ε from a reference vector w ₁ that already exists, the reference vector w ₁ and its unit and a value greater than zero. The reference vectors w ₂ and w _{3 of the} units directly coupled with the coupling strength of are updated so as to approach the learning vector xa. Further, when the learning vector xb is input at a position away from the existing reference vectors w _{1 to} w ₆ by a predetermined distance ε, a reference vector (not shown) is newly added at the position of the learning vector xb.

That is, in this data processing apparatus, when the learning vector x is input, it is checked whether or not the reference vector w _i satisfying the following expression (3) already exists.

When such a reference vector w _i exists, the unit directly coupled with the reference vector w _i and the unit with a coupling strength greater than zero according to the following equations (4) to (7): Update the reference vector of. If no link exists between the reference vector closest to the currently input learning vector and the next closest reference vector (second nearest neighbor reference vector), a new link is generated. Then, the coupling strength is determined according to the following equation (9), and the second nearest neighbor reference vector is updated.

In Equation (5), γa and γb are learning coefficients when a learning vector x having a positive evaluation is input and a learning vector x having a negative evaluation is input, respectively. σ _γ is a parameter that determines the degree to which the update amount is reduced according to the distance between both vectors when the reference vector is separated from the currently input learning vector x. E is an evaluation value for the learning vector, and is a positive value when the evaluation is positive, and a negative value when the evaluation is negative. Ω (b) is defined by equation (7), and b is the number of the unit (nearest neighbor unit) that is closest to the learning vector (here, Euclidean distance). s (i, j) represents the coupling strength between the i-th unit and the j-th unit. That is, Ω (b) represents a set of numbers of the nearest unit and the unit (neighboring unit) coupled to the nearest unit with a coupling strength having a value greater than zero.

Further, a _i is defined by a function of decreasing the distance between the reference vector w _i of the i-th unit and the learning vector x as shown in Expression (6), and the i-th unit is determined from the learning vector x. The degree of influence, that is, the degree of influence is shown.

Then, the reference vector w _i of the i-th unit is updated according to the above equation if the i-th unit is the nearest unit or its neighboring unit, as shown in Equation (5). If there is no update. The update amount Δw _i in the case of updating is proportional to the influence degree a _i from the learning vector x and the evaluation value E, as shown in Expression (5). For this reason, the larger the influence degree a _{i and the} larger the absolute value of the evaluation value E, the larger the reference vector w _i approaches or leaves the learning vector x. When a learning vector x having a positive evaluation value E is input, the nearest neighbor unit and the reference vector w _i of the neighboring unit move in the reference vector space toward the position of the learning vector x. On the other hand, when a learning vector x having a negative evaluation value E is input, the nearest neighbor unit and the reference vector w _i of the neighboring unit move in the reference vector space in a direction away from the learning vector x.

Further, when the learning vector x is input, if the reference vector w _i that satisfies the above equation (3) does not yet exist and the evaluation value E of the input learning vector x exceeds a predetermined threshold value. Then, according to the following equation (8), the (N + 1) th unit having the same reference vector wN _{+ 1} as the learning vector is newly generated on the competitive layer.

Further, in this data processing apparatus, the coupling strength s (i, j) is stored for each pair of units arranged in the competitive layer, the learning vector x is input, and the nearest neighbor unit and its neighboring units are referred to. When the vector w _i is updated, the coupling strength s (i, j) is updated for these units according to the following equation (9). Here, β is a parameter indicating the update amount. As shown in the equation (9), the coupling strength s (i, j) between the i-th unit and the j-th unit increases when the evaluation value E is positive, and decreases when it is negative. At this time, the greater the degree of influence a _i and a _j of each unit from the learning vector x, the greater the change.

  Further, when the learning vector x is inputted, if the nearest unit does not exist, a new unit is added to the competitive layer as described above, and the added unit and two neighboring units are Are newly calculated according to the above equation (9).

  Further, in this data processing apparatus, the activity is managed for each unit, and the activity Sa (i) of the i-th unit is updated by the following equation (10). The initial value of the activity Sa (i) may be set as appropriate.

When the learning vector x is input and the nearest neighbor unit and the reference vector w _{i of the} neighboring unit are updated, the degree of activity a _i and the degree of influence a _i and An amount proportional to the evaluation value E is added. For this reason, when a learning vector x having a positive evaluation value is input in the vicinity, the activity Sa (i) increases, and conversely, when a learning vector x having a negative evaluation value is input in the vicinity, the activity Sa (i) is increased. ) Goes down. In this data processing device, when the activity Sa (i) of the i-th unit becomes less than a predetermined value, the unit is deleted from the competitive layer. That is, the reference vector w _i , the activity Sa (i), and the coupling strength related to the unit are all erased.

By repeating the learning process described above, the reference vector group shows the relationship between the input data u and the output data v of preferable evaluation, as shown in FIG. In this data processing apparatus, when the input data p is input in the execution mode, the similarity z _i between the input data p and the component u _i of the input data of each reference vector w _i is calculated according to the following equation (11). The preferred output data q corresponding to the input data p is calculated by performing weighted addition of the component v _i of the output data of the reference vector w _i using the similarity z _i according to the following equation (12). To do. Here, Σ is the sum of a set of units that are directly or indirectly made by a coupling strength s (i, j) greater than zero. FIG. 3 schematically shows the state of this calculation.

  FIG. 4 is a functional block diagram of a data processing apparatus according to an embodiment of the present invention. The data processing apparatus is configured to include a known computer system, and the functions shown in the figure are realized by executing a program to which the present invention is applied by the computer system. This program may be installed in a computer by a computer-readable information storage medium such as a CD-ROM or a DVD-ROM, or may be downloaded from a communication network such as the Internet.

  As shown in the figure, the data processing device functionally includes a learning vector input unit 10, a distance calculation unit 12, a reference vector update unit 14, a reference vector addition unit 16, a data output unit 18, a reference vector / activity level storage. A unit 20, a coupling strength storage unit 22, and a reference vector / coupling strength deletion unit 24.

First, the reference vector / activity level storage unit 20 stores the reference vector of each unit arranged in the competitive layer, and stores the activity level of each unit in association with the reference vector. FIG. 5 schematically shows the storage contents of the reference vector / activity storage unit 20, and as shown in FIG. 5, the input included in the reference vector w _{i of the} unit in association with the number i of each unit. Data u _i, output data v _i , and activity Sa (i) are stored.

  The coupling strength storage unit 22 stores the coupling strength S (i, j) of the pair in association with each pair of reference vectors stored in the reference vector / activity storage unit 20.

  The learning vector input unit 10 inputs a large number of learning vectors x composed of input data p and output data q in order.

The distance calculation unit 12 calculates a distance (for example, Euclidean distance) between the learning vector x sequentially input and each reference vector w _i stored in the reference vector / activity storage unit 20. Then, it is determined whether or not the minimum value of the distance is equal to or smaller than the predetermined distance ε, that is, whether or not there is a unit that satisfies the expression (3). If such a unit exists, the reference vector update unit 14 is instructed, and if not, the reference vector addition unit 16 is instructed to perform the processing.

When the minimum value of the distance calculated by the distance calculation unit 12 is larger than the predetermined distance ε, the reference vector adding unit 16 uses the reference vector w _{N + 1} determined by the equation (8) based on the learning vector x as the reference vector It is additionally stored in the activity storage unit 20. Further, the activity degree Sa (N + 1) of the (N + 1) th unit is also stored in association with the reference vector wN _{+ 1} . The value of the activity Sa (N + 1) is calculated according to the equation (10).

  Further, the reference vector adding unit 16 has the added N + 1th unit, the unit closest to the unit (the unit closest to the distance between the reference vectors), and the next closest unit (the distance between the reference vectors is The bond strength with the next closest unit is calculated according to equation (9) and stored in the bond strength storage unit 22.

When the minimum distance calculated by the distance calculation unit 12 is equal to or less than the predetermined distance ε, the reference vector update unit 14 converts a part of the reference vector stored in the reference vector / activity storage unit 20 to Equation (4). ) To Equation (7), the learning vector x is moved closer to or away from the learning vector x according to the evaluation value E. First, when there is no link between the nearest reference vector of the current learning vector x and the second nearest neighbor reference vector, a link is generated. Thereafter, the nearest unit and the reference vector of the neighborhood unit are moved closer to or away from the learning vector x. Further, the activity Sa (i) of the i-th unit related to the updated reference vector w _i is updated according to the equation (10). At this time, as the distance between the updated reference vector w _i and the learning vector x is shorter, the activity sa (i) of the i-th unit related to the updated reference vector w _i is greatly changed. Also, if a learning vector x having a positive evaluation value E is input, the activity Sa (i) is increased, and if a learning vector x having a negative evaluation value E is input, the activity Sa (i). Decrease.

  Further, the reference vector update unit 14 updates the coupling strength s (i, j) according to the equation (9) for the pair of the nearest unit and the unit for which the unit and the link are set.

Referring vector coupling strength deleting unit 24, when associated with a reference vector w _i activity Sa (i) is lower than a predetermined threshold value, the reference vector w _i and the activity Sa to its associated with (i) Delete from the reference vector / activity level storage unit 20. At the same time, the connection strength related to the unit related to the deleted reference vector w _i is deleted from the connection strength storage unit 22.

  The data output unit 18 operates exclusively in the execution mode, receives the input data p, calculates the output data q based on the equations (11) and (12), and outputs it.

FIG. 6 is a flowchart showing learning processing of the data processing apparatus according to the present embodiment. As shown in the figure, when a learning vector x is input (S101), it is determined whether a unit is arranged in the competitive layer, that is, whether one or more records exist in the reference vector / activity storage unit 20. (S102). If it does not exist, the reference vector w ₁ identical to the learning vector x and the activity Sa (1) calculated by the equation (10) are stored in the reference vector / activity storage unit 20 (S103). On the other hand, if it exists, the distance calculation unit 12 searches for the nearest unit of the learning vector x (S104). Then, it is determined whether the distance between the reference vector of the nearest unit and the learning vector is greater than a predetermined distance ε (S105). If the evaluation value E of the current learning vector x is greater than the predetermined distance ε and exceeds a predetermined threshold value (S106), the reference vector adding unit 16 uses the same reference vector w _{N + 1} and expression as the learning vector x. The activity Sa (N + 1) calculated by (10) is stored in the reference vector / activity storage unit 20 (S107). Further, the coupling strength between the added (N + 1) th unit and two units close to the unit is calculated and stored in the coupling strength storage unit 22 (S108).

  On the other hand, when it is determined in S105 that the distance between the learning vector x and the reference vector of the nearest neighbor unit is equal to or less than the predetermined distance ε, the reference vector update unit 14 refers to the reference vector for the nearest neighbor unit and the neighborhood unit. Is updated according to the equations (4) to (7). Further, the coupling strength of the link connected to the nearest unit is updated according to the equation (9). At this time, if no link exists between the nearest neighbor reference vector and the second nearest neighbor reference vector, a link is generated, and then the update process is executed (S109).

  Next, the nearest neighbor unit and the activity of the neighbor unit are updated according to equation (10). If the updated activity level is lower than the user-specified threshold, the reference vectors, activity levels, and coupling strengths for these units are deleted from the reference vector / coupling level storage unit 20 and the coupling strength storage unit 22 (S110). ).

  Thereafter, it is determined whether the processes of S101 to S110 have been repeated as many times as specified by the user (S111). If it has been repeated, the minimum one of the bond strengths stored in the bond strength storage unit 22 is deleted, and the count of the number of repetitions is reset (S112). That is, the weakest link is deleted every time the learning vector is input a predetermined number of times. Thereafter, it is determined whether or not the learning process has been completed for all the learning data (S113), and the processes after S101 are repeatedly executed until the learning process is completed.

  According to the above embodiment, the activity is stored in association with each reference vector. This degree of activity is updated based on the distance between the learning vector and the reference vector and the evaluation with respect to the learning vector. In the present embodiment, reference vectors and the like are deleted from the reference vector / activity storage unit 20 in accordance with the degrees of activity stored / updated in this way, so that reference vectors that are not based on the properties of the currently input learning vectors The problem of being stored in the reference vector / activity level storage unit 20 can be prevented. Therefore, even if the properties of the learning vectors input sequentially change with time, the intended vector set can be suitably reconstructed as a reference vector group.

It is a figure explaining the update of a reference vector. It is a figure which shows the learned reference vector group. It is a figure explaining the process in execution mode. It is a functional block diagram of the data processor concerning this embodiment. It is a figure which shows the memory content of a reference vector and activity memory | storage part. It is a flowchart which shows the learning process of a data processor.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 Learning vector input part, 12 Distance calculation part, 14 Reference vector update part, 16 Reference vector addition part, 18 Data output part, 20 Reference vector and activity memory | storage part, 22 Bond strength memory | storage part, 24 Reference vector and bond strength deletion Department.

Claims (8)

Reference vector storage means for storing one or more reference vectors;
Distance calculating means for calculating the distance between the learning vector sequentially input and each reference vector stored in the reference vector storage means;
Reference vector addition means for adding a reference vector determined based on the learning vector to the reference vector storage means and storing it when the minimum distance calculated by the distance calculation means is larger than a predetermined value;
When the minimum distance calculated by the distance calculation means is less than or equal to a predetermined value, at least a part of the one or more reference vectors stored in the reference vector storage means is set to the learning vector with respect to the learning vector. Reference vector update means for approaching or separating according to the evaluation of
Activity is stored in association with each reference vector stored in the reference vector storage means, and the activity is updated based on a distance between the reference vector and the learning vector and evaluation of the learning vector. Memory update means,
Reference vector deleting means for deleting each reference vector from the reference vector storage means in accordance with the degree of activity associated with each reference vector;
A data processing apparatus comprising: The data processing apparatus according to claim 1,
The reference vector adding means stores the reference vector in the reference vector storage when the minimum value of the distance calculated by the distance calculating means is larger than the predetermined value and the evaluation with respect to the learning vector is not less than a predetermined value. Add to the means and remember,
A data processing apparatus. The data processing apparatus according to claim 1 or 2,
The activity storage update means changes the activity greatly as the distance between the reference vector and the learning vector is shorter.
A data processing apparatus. The data processing apparatus according to any one of claims 1 to 3,
The activity memory update means increases the activity when the evaluation with respect to the learning vector is positive, and decreases the activity when the evaluation is negative.
A data processing apparatus. The data processing device according to any one of claims 1 to 4,
A coupling that stores the coupling strength of the pair in association with each pair of the reference vectors stored in the reference vector storage means and updates the coupling strength according to the distance between each of the pair and the learning vector. A strength memory updating means;
A data processing apparatus. The data processing apparatus according to claim 5, wherein
The reference vector update means causes the reference vector related to the minimum value and a reference vector in the vicinity of the reference vector determined based on the coupling strength to be close to or separated from the learning vector,
The reference vector deletion means, when deleting the reference vector from the reference vector storage means, deletes the joint strength associated with the reference vector together;
A data processing apparatus. A distance calculating step of calculating a distance between the learning vector sequentially input and each of the reference vectors stored in the reference vector storage unit that stores one or more reference vectors;
A reference vector adding step of adding and storing a reference vector determined based on the learning vector in the reference vector storage means when a minimum value of the distance calculated in the distance calculating step is larger than a predetermined value;
When the minimum value of the distance calculated in the distance calculating step is less than or equal to a predetermined value, at least a part of the one or more reference vectors stored in the reference vector storage unit is set to the learning vector. A reference vector update step that approaches or separates depending on the evaluation of
An activity update step of updating the activity of the reference vector based on the distance between each reference vector stored in the reference vector storage means and the learning vector and the evaluation of the learning vector;
A reference vector deleting step of deleting each reference vector from the reference vector storage unit according to the activity of each reference vector;
A data processing method comprising: Reference vector storage means for storing one or more reference vectors;
Distance calculation means for calculating the distance between the learning vector sequentially input and each reference vector stored in the reference vector storage means;
Reference vector addition means for adding a reference vector determined based on the learning vector to the reference vector storage means and storing it when the minimum value of the distance calculated by the distance calculation means is larger than a predetermined value;
When the minimum distance calculated by the distance calculation means is less than or equal to a predetermined value, at least a part of the one or more reference vectors stored in the reference vector storage means is set to the learning vector with respect to the learning vector. Reference vector updating means for approaching or separating according to the evaluation of
Activity is stored in association with each reference vector stored in the reference vector storage means, and the activity is updated based on a distance between the reference vector and the learning vector and evaluation of the learning vector. A program for causing a computer to function as a reference vector deleting unit that deletes each reference vector from the reference vector storage unit according to the degree of activity associated with each reference vector.

Images