JP2007280008A - Data processor, data processing method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To allow a robot or the like to perform a task requesting real time performance. <P>SOLUTION: Input networks net<SB>1</SB>and net<SB>3</SB>and an output network net<SB>2</SB>composed of a plurality of nodes having time-series pattern models expressing time-series patterns are updated like self-organization so that the connection weight of winner nodes for input data of frames at a time t of the input network net<SB>1</SB>and winner nodes for the input data of frames at a time t' delayed only by a fixed time from the time t of the input network net<SB>3</SB>can be strengthened, and so that the connection weight of winner nodes for the input data of frames at the time t' of the input network net<SB>3</SB>and winner nodes for the output data of frames at the time t' of the output network net<SB>2</SB>can be strengthened. This invention may be applied to, for example, the robot or the like. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a data processing apparatus, a data processing method, and a program, and more particularly, for example, a data processing apparatus, a data processing method, and a program that enable a robot or the like to perform a task that requires real-time performance. About.

  For example, a forward model or an inverse model can be used to realize a robot that performs tasks autonomously.

  FIG. 1 shows the concept of a forward model and an inverse model.

  Control data that outputs output data as other time-series data is given to input data as certain time-series data (time-series data), but detailed information about the control object is not known (inside the control object) However, it is assumed that the input data given to the control object and the output data obtained from the control object with respect to the input data can be observed.

  Here, the input data given to the control object and the output data obtained from the control object for the input data may be any physical quantity as long as it can be observed. The control target may be any target (thing) as long as input data can be given and output data can be obtained for the input data.

  Therefore, for example, a ball, a musical instrument, an automobile, a gas heater, and other various objects can be controlled. That is, for example, by applying (applying) force as input data for the ball, the position and speed of the ball as output data that changes with respect to the input data can be obtained. Also, for example, for a car, by operating (giving) a steering wheel, an accelerator, a brake, etc. as input data, the position and speed of the car as output data that changes with respect to the input data can be obtained it can. Further, for example, for a gas heater, the temperature of the room as output data that changes with respect to the input data can be obtained by adjusting the heating power as input data.

  As described above, when input data is given to a control target and output data is obtained for the input data, a model (model) of the control target is a forward model.

  In the forward model, when input data is input (when input data is given), an estimated value of output data obtained from the control target is output for the input data. Therefore, according to the forward model, output data obtained from the control target can be estimated for the input data without giving actual input data to the control target.

  On the other hand, in order to determine the target value of the output data obtained from the controlled object and obtain the output data that becomes the target value, a model that can estimate the input data to be given to the controlled object is an inverse model. The forward model can be regarded as a mapping from input data to output data, but its inverse mapping is an inverse model.

  Here, in order to obtain output data that is a target value, which is obtained by the inverse model, the input data to be given to the control target is hereinafter also referred to as control data as appropriate.

  The forward model and the inverse model as described above can be used for the robot (configuration) as described above.

  That is, the robot now has a microphone (microphone) and camera, can input voice (sound) data and image data, and has a speaker and an actuator (motor) to output voice (voice data). The arm can be moved by driving the motor according to the motor data (motor signal).

  In such a robot, when outputting voice data as output data according to input data such as certain voice data or image data, and outputting motor data as output data for operating a desired arm, conventionally, Is what voice data should be output according to the recognition result of recognizing voice data and image data input to the robot using a voice recognition device or image recognition device, or what motor data Is to be programmed (designed) in advance.

  On the other hand, if a forward model is used, as shown in FIG. 2, a robot that outputs desired voice data or motor data for performing a desired arm operation according to certain voice data or image data is controlled. It is possible to configure an actual robot as a forward model of a robot assumed as a control target (hereinafter referred to as an assumed robot as appropriate). That is, if the actual robot can learn the relationship between the input data and the output data for the assumed robot, a robot as a forward model of the assumed robot can be configured.

  Specifically, a set of input data such as voice data and image data to be input to the assumed robot and output data such as voice data and motor data to be output by the assumed robot corresponding to each input data is prepared in advance. Give to a real robot. In an actual robot, a forward model of an assumed robot that estimates (outputs) output data corresponding to input data using only a set of input data and output data given from outside (hereinafter referred to as teaching data as appropriate). Can be obtained, output data such as desired voice data and motor data can be output in accordance with input data such as voice data and image data actually input.

  Further, if the inverse model is used, as shown in FIG. 3, it is possible to configure an arm controller that controls the arm of the robot as a control target and controls the arm that is the control target.

  That is, it is assumed that the arm of the robot is moved by a motor driven in accordance with motor data as input data, and as a result, the position of the tip of the arm changes. Further, based on a three-dimensional coordinate system in which the center of gravity of the robot is the origin, the forward (front) direction of the robot is the x axis, the right direction (viewed from the robot) is the y axis, and the upward direction is the z axis. The position of the tip of the arm is represented by the (x, y, z) coordinates of the three-dimensional coordinate system. In this case, the motor is driven according to the motor data, and the position of the tip of the arm is changed to draw a locus with the tip of the arm. Here, a sequence of coordinates of a locus (tip position locus) drawn by the tip of the arm is referred to as tip position locus data.

  In order for the arm to draw a desired tip position locus, that is, to output the desired tip position locus data as output data, the arm is drawn so that the arm draws such a tip position locus. The motor data to be driven needs to be given to the motor as input data.

  Now, using only teaching data that is a set of motor data as input data and tip position locus data as output data when the motor data is given to the motor, a tip position locus as output data If the inverse model of the arm that estimates the motor data as input data (control data) that can obtain the data as the target value can be obtained, the inverse model corresponds to the tip position locus data that is the target value. It can be used for an arm controller for determining motor data to be performed.

  According to the arm controller as an inverse model of such an arm, when the tip position trajectory data as input data is input to the robot, the robot uses the arm controller, and the corresponding motor data (control Data) can be determined. When the robot drives the motor according to the motor data, the robot arm moves so as to draw a locus corresponding to the tip position locus data as the input data.

  As described above, if the forward model and inverse model can be obtained using only the set of input data and output data (teaching data), the output corresponding to each input data can be obtained using the forward model and inverse model. A robot that outputs data can be easily configured.

  As a method for obtaining the forward model and the inverse model as described above, there is modeling using a linear system.

  In modeling using a linear system, for example, as shown in FIG. 4, the input data to the controlled object at time t is u (t), the output data is y (t), and the output data y ( The relationship between t) and input data u (t), that is, the controlled object is approximated as a linear system given by, for example, Expression (1) and Expression (2).

  Here, x (t) is called a state variable of the linear system at time t, and A, B, and C are coefficients. Further, here, for the sake of simplicity, it is assumed that the input data u (t) and the output data y (t) are one-dimensional vectors (scalars) and the state variable x (t) is an n-dimensional vector ( Here, n is an integer value of 2 or more), and A, B, and C are constant matrices given by an n × n matrix, an n × 1 matrix, and a 1 × n matrix, respectively.

  In modeling using a linear system, the relationship between the input data u (t) that can be observed and the output data y (t) that is observed when the input data u (t) is given to the controlled object However, by determining the row examples A, B, and C so as to satisfy the equations (1) and (2), the forward model of the controlled object is obtained.

  However, modeling using a linear system is not sufficient to model a complicated control object, for example, a control object having nonlinear characteristics.

  In other words, the actual controlled object is complex and often has non-linear characteristics. However, when this controlled object is approximated and modeled as a simple linear system, the output that the forward model estimates for the input data An estimation error of data or input data (control data) estimated by the inverse model with respect to output data becomes large, and it is difficult to perform highly accurate estimation.

  Therefore, as a method for obtaining a forward model or an inverse model for a controlled object having nonlinear characteristics, for example, using a neural network, teaching data, that is, input data given to the controlled object and its input data are given. There is a method of learning a set with output data observed from a controlled object sometimes. Here, a neural network is a network configured by interconnecting artificial elements simulating living nerve cells (neurons), and the relationship between teaching data given from outside, that is, input data and output data. You can learn the relationship.

  However, in order to appropriately model a control target with a neural network, it is necessary to increase the scale of the neural network in accordance with the complexity of the control target. As the scale of the neural network increases, the time required for learning increases dramatically, and stable learning becomes difficult. This is the same when the number of dimensions of input data or output data is large.

  On the other hand, when a forward model or inverse model is obtained using only a set of input data and output data (teaching data), learning is performed using the teaching data, and the teaching data is one of several patterns. It is necessary to recognize which is the case. That is, it is necessary to learn and recognize patterns of input data and output data as teaching data.

  The technique of learning and recognizing patterns is generally called pattern recognition, and learning in pattern recognition can be divided into supervised learning and unsupervised learning. .

  Supervised learning is a method in which learning data belonging to each pattern is given information (this is called a correct answer label) and learning data belonging to that pattern is learned for each pattern. Many learning methods using networks and HMM (Hidden Markov Model) have been proposed.

  Here, FIG. 5 shows an example of supervised learning.

  In supervised learning, learning data used for learning is prepared for each assumed category (class) (for example, each phoneme category, each phoneme category, each word category, etc.). For example, when learning speech data of utterances “A”, “B”, and “C”, a large number of speech data of “A”, “B”, and “C” are prepared.

  On the other hand, a model used for learning (a model for learning learning data of each category) is also prepared for each assumed category. Here, the model is defined by parameters. For example, for learning speech data, an HMM or the like is used as a model. The HMM is a state transition probability of transition from one state to another state (including the original state) or an observation value output from the HMM. It is defined by an output probability density function representing the probability density of

  In supervised learning, learning of a model of each category (class) is performed using only the learning data of that category. That is, in FIG. 5, the model “A” learning is performed using only the category “A” learning data, and the category “B” model learning is performed using only the category “B” learning data. Done. Similarly, learning of the model of category “C” is performed using only the learning data of category “C”.

  In supervised learning, it is necessary to learn the model of the category using the learning data of each category as described above. Therefore, learning data of the category is prepared for each category, and the model of each category is prepared. For learning, learning data of the category is given, and a model for each category is obtained. As a result, according to supervised learning, a template for each class (class (category) model represented by the correct answer label) can be obtained based on the correct answer label.

  At the time of recognition, for a certain recognition target data, a template (the template with the highest likelihood) that matches the recognition target data is obtained, and the correct answer label of the template is output as a recognition result. The

  On the other hand, unsupervised learning is learning that is performed in a situation where the correct label is not given to the learning data of each pattern. For example, there is a learning method using a neural network or the like, but the correct answer label is not given. This is very different from supervised learning.

  By the way, pattern recognition can be regarded as quantization of a signal space in which data (signals) to be recognized to be recognized by the pattern recognition is observed. In particular, pattern recognition when the recognition target data is a vector may be referred to as vector quantization.

  In vector quantization learning (codebook generation), a representative vector corresponding to a class (referred to as a centroid vector) is arranged in a signal space where data to be recognized is observed.

  One of the typical techniques for unsupervised learning of vector quantization is the K-means clustering method. In the K-means method, as an initial state, a centroid vector is appropriately arranged, a vector as learning data is assigned to the nearest centroid vector, and an average vector of learning data assigned to each centroid vector is determined. This is a learning method that repeats updating the centroid vector. A collection of centroid vectors is called a code book.

  Here, a method of accumulating a large number of learning data and performing learning using all of them is called batch learning, and the K-means method is classified as batch learning. For batch learning, when learning data is observed, learning is performed using the learning data, and the parameters (centroid vector components, output probability density function that defines the HMM, etc.) are updated little by little. , Called on-line learning.

  As online learning, learning by SOM (self-organization map) proposed by T. Kohonen is known. In learning by SOM, the connection weight of the input layer and output layer of SOM is updated (corrected) little by little by online learning.

  That is, in the SOM, the output layer has a plurality of nodes, and a weight vector is given to each node of the output layer. When this weight vector is a centroid vector, learning in vector quantization can be performed.

  Specifically, in the node of the output layer of the SOM, the node having the closest distance between the weight vector and the vector as the learning data is determined as the winner node that best matches the vector as the learning data, and the winner node Are updated so as to approach the vector as learning data. Furthermore, the weight vector of the node in the vicinity of the winner node is also updated so as to be slightly closer to the learning data. As a result, as learning progresses, similar nodes are arranged on the output layer so that similar nodes become closer and dissimilar nodes become farther away. Therefore, a map corresponding to the pattern included in the learning data is configured in the output layer. In this way, as learning progresses, similar nodes (nodes with similar weight vectors) gather near each other and a map corresponding to the pattern included in the learning data is constructed by self-organization or self This is called self-organization.

  Here, in the K-means method, only the vector having the closest distance to the learning data is updated, so the updating method is called WTA (winner-take-all). On the other hand, in learning by SOM, not only the weight vector of the node closest to the learning data (winner node) but also the weight vector of the node near the winner node is updated, so the update method is SMA (soft-max called adaptation). It is known that when learning with WTA, the learning result tends to fall into a local solution, whereas when learning with SMA, the problem of falling into a local solution can be improved.

  Note that SOM is described in Non-Patent Document 1, for example.

T. Kohonen, “Self-Organizing Map”, Springer Fairlark Tokyo

  By the way, in order to make the behavior (behavior) of the robot in the real world more natural, research is being conducted on a framework for the robot itself to acquire the structure of cognitive behavior through its own behavior. This framework is called “Cognitive Behavioral Architecture”. Here, the recognition behavior means that, for example, a robot or the like recognizes (recognizes) an external state (including the state of the robot itself) and takes an action according to the recognition result.

  While many of the previous robot behavior control methodologies have been pre-made (programming by robot designers, etc.) to express their behavior, the “cognitive behavior architecture” is made as much as possible. Is a methodology for acquiring cognitive behavior based on self-search, teaching, and imitation. A robot that has acquired an action with such a “cognitive behavior architecture” may adapt to a given environment and behave robustly.

  The “cognitive behavior architecture” is abstracted into a problem of calculating appropriate motor data as motor data supplied to a motor that drives a robot, for example, with respect to sensor data output by a sensor that detects an external state. be able to. “Appropriate” here has a deep meaning of being intelligent.

  In general, both sensor data output from a sensor and motor data supplied to a motor are continuous time-series data. Further, the action to be taken now needs to be determined in consideration of not only the current external state but also the past external state, the past cognitive behavior, and the like as necessary. In other words, there is a need to handle context in determining the cognitive behavior that should be taken now. Furthermore, a robot that performs cognitive behavior in the real world needs to handle data with a large number of dimensions as sensor data or motor data. In addition, the behavior of sensor data and motor data handled by such robots is complicated, and it is difficult to model with a linear system. In order to apply the “cognitive behavior architecture” to the real world, a method for efficiently handling the sensor data and motor data as described above is required.

  Therefore, a time-series pattern storage network composed of a plurality of nodes having a time-series pattern model that represents a time-series pattern that is a pattern of time-series data such as sensor data and motor data, and a time series of multidimensional vectors When learning time-series data such as sensor data and motor data in a self-organized manner, further learning time-series data as input data, and learning time-series data as output data Regarding the method of combining nodes with a sequence pattern storage network, recognizing an external state based on input data, and generating output data corresponding to the action to be taken by the robot according to the recognition result, Have previously proposed (for example, Japanese Patent Application No. 2004-353382).

  According to the previously proposed method, the correspondence between input data and output data, such as a time series of multidimensional vectors (for example, imitating 50-sound audio data and its audio data) ) Correspondence relationship with audio data) and cross-modal correspondence relationships such as correspondence relationship between sensor data and motor data (for example, information acquired by a person through vision (vision) and actions taken on the information ( The correspondence relationship with the exercise to be performed) can be learned.

  However, in the method proposed earlier, self-organized structure generation of time-series data and static correspondence between time-series data are learned. As tasks that can be applied to this method, May be limited to tasks that do not require real-time capability.

  That is, for example, in a task of chasing an object that is moving, the task of recognizing the state of the object that is actually moving and the action of chasing the object according to the recognition result are implemented. Although it is required to perform it continuously in time, it is sometimes difficult to learn such a task in the method proposed above.

  The present invention has been made in view of such a situation. For example, a robot or the like can perform a task that requires real-time performance.

  A data processing device according to a first aspect of the present invention is a data processing device that learns a relationship between input data that is time-series data and output data that is other time-series data. Input / output data extracting means for extracting input data in time units and extracting the output data in predetermined time units from the output data, and expressing a time series pattern that is a pattern of time series data as the input data A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model, a second input time series pattern storage network, and a time series pattern which is a pattern of time series data as the output data Output time series pattern memory composed of multiple nodes with time series pattern model to express A second input time-series pattern storage network and a node of the first input time-series pattern storage network and a node of the second input time-series pattern storage network. Among the nodes of the first input time-series pattern storage network in the input / output relation model in which the nodes of the output time-series pattern storage network are coupled to each other. A first recognition learning process for determining a first input winner node that is a suitable node and updating the first input time-series pattern storage network in a self-organized manner based on the first input winner node And a predetermined time unit input data from the second input time-series pattern storage network node. A second input winner node that is a node that best fits the data, and based on the second input winner node, the second input time-series pattern storage network is updated in a self-organizing manner. From the recognition learning processing means and the node of the output time series pattern storage network, determine the output winner node that is the most suitable node for the output data of the predetermined time unit, based on the output winner node, A generation learning processing means for updating the output time-series pattern storage network in a self-organized manner, the first input winner node for the input data in the predetermined time unit of the first input time-series pattern storage network, The predetermined time delayed by a predetermined time from the time of the input data in the predetermined time unit of the second input time series pattern storage network A process for updating the second input time series pattern storage network to update the connection weight representing the degree of connection with the second input winner node with respect to the input data in the time unit, and the input in the predetermined time unit; The second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the data and the time of the input data of the predetermined time unit of the output time series pattern storage network A connection weight updating unit that performs a process of updating the connection data representing the degree of connection with the output winner node with respect to the output data of the predetermined time unit delayed by the time.

  A data processing method or program according to the first aspect of the present invention is a data processing method for a data processing apparatus that learns a relationship between input data that is time-series data and output data that is other time-series data, Alternatively, in a program that causes a computer to execute data processing for learning the relationship between input data that is time-series data and output data that is other time-series data, input data in a predetermined time unit is input from the input data. It is composed of a plurality of nodes having a time-series pattern model that extracts the output data of the predetermined time unit from the output data and expresses a time-series pattern that is a pattern of time-series data as the input data. A first input time-series pattern storage network and a second input time-series pattern storage network, An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as output data, and the first input time series pattern A node of the storage network and a node of the second input time-series pattern storage network, and a node of the second input time-series pattern storage network; a node of the output time-series pattern storage network; A first input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the first input time-series pattern storage network in the input / output relationship model to which , Based on the first input winner node, the first input time series pattern description The network is updated in a self-organized manner, and a second input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the second input time series pattern storage network. The second input time-series pattern storage network is updated in a self-organized manner based on the second input winner node, and the predetermined time unit is updated from the nodes of the output time-series pattern storage network. An output winner node which is a node most suitable for output data is determined, and based on the output winner node, the output time series pattern storage network is updated in a self-organized manner, and the first input time series pattern storage network is updated. The first input winner node for the input data of the predetermined time unit and the second input time series pattern storage network The work weight is updated so as to increase the connection weight indicating the degree of connection with the second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit. And a second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the second input time-series pattern storage network A connection weight representing a degree of connection with the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the output time series pattern storage network. And a step of performing a process of updating to strengthen.

  In the first aspect of the present invention as described above, input data in a predetermined time unit is extracted from the input data, and output data in the predetermined time unit is extracted from the output data. Further, a first input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the first input time-series pattern storage network, and the first input winner node is determined. The first input time-series pattern storage network is updated in a self-organized manner, and the input data of the predetermined time unit is the most out of the nodes of the second input time-series pattern storage network. Determining a second input winner node that is a matching node, updating the second input time-series pattern storage network in a self-organized manner based on the second input winner node, and the output From among the nodes of the time-series pattern storage network, an output winner node that is a node most suitable for the output data of the predetermined time unit is determined. Based on the output winning node, the output time series pattern storage network, it is performed to update self-organizing manner. The first input winner node for the input data of the predetermined time unit of the first input time series pattern storage network and the input of the predetermined time unit of the second input time series pattern storage network A connection weight representing the degree of connection with the second input winner node for the input data in the predetermined time unit delayed by a certain time from the time of the data is updated so as to increase, and at the time of the second input The second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the sequence pattern storage network, and the output time series pattern storage network of The output data of the predetermined time unit delayed by a certain time from the time of the input data of the predetermined time unit Connection weight representing the degree of coupling between the output winning node is updated to enhance.

  A data processing device according to a second aspect of the present invention is a data processing device that generates output data that is other time-series data with respect to input data that is time-series data. Input data extracting means for extracting input data, and a first input time series pattern storage composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data Network, second input time-series pattern storage network, and output time-series pattern storage composed of a plurality of nodes having a time-series pattern model representing a time-series pattern that is a pattern of time-series data as the output data A node of the first input time-series pattern storage network, A node of the second input time series pattern storage network is coupled, and a node of the second input time series pattern storage network and a node of the output time series pattern storage network are coupled. An input winner node determining means for determining an input winner node that is a node most suitable for the input data of the predetermined time unit from among the nodes of the first input time-series pattern storage network in the output relation model; Finding the strongest coupling node that is the strongest node with the input winner node from among the two input time-series pattern storage network nodes, and selecting the strongest coupling node from among the nodes of the output time-series pattern storage network The node that has the strongest connection with is determined as the generation node that generates the output data And generating node determining means that, based on the time series pattern models the generation node has, and a generation means for generating the output data.

  A data processing method or program according to a second aspect of the present invention provides a data processing method or a time series for a data processing device that generates output data that is other time-series data for input data that is time-series data. In a program that causes a computer to execute data processing for generating output data that is other time-series data with respect to input data that is data, input data in a predetermined time unit is extracted from the input data, and the input data A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern that is a pattern of time series data as a second input time series pattern storage network, A time series expressing a time series pattern which is a pattern of time series data as the output data An output time-series pattern storage network composed of a plurality of nodes having a turn model, and a node of the first input time-series pattern storage network and a node of the second input time-series pattern storage network The first input time series pattern storage network in the input / output relation model in which the nodes of the second input time series pattern storage network and the nodes of the output time series pattern storage network are coupled. An input winner node which is a node most suitable for the input data of the predetermined time unit is determined from among the nodes of the second input time series pattern storage network, and the input winner node is determined from the nodes of the second input time series pattern storage network Find the strongest coupling node that is the strongest node of coupling, and output the time series pattern A node having the strongest connection with the strongest connection node is determined as a generation node for generating the output data from the nodes of the storage network, and the output data is based on the time-series pattern model of the generation node. The step of generating is included.

  In the second aspect of the present invention as described above, input data in a predetermined time unit is extracted from the input data, and the predetermined time unit is extracted from the nodes of the first input time series pattern storage network. The input winner node which is the node most suitable for the input data is determined. Further, a strongest connection node that is a node having the strongest connection with the input winner node is obtained from the nodes of the second input time-series pattern storage network, and is selected from the nodes of the output time-series pattern storage network. The node having the strongest connection with the strongest connection node is determined as the generation node that generates the output data. Then, the output data is generated based on the time series pattern model of the generation node.

  A data processing device according to a third aspect of the present invention is a data processing device that learns a relationship between input data that is time-series data and output data that is other time-series data. Input / output data extracting means for extracting input data in time units and extracting the output data in predetermined time units from the output data, and expressing a time series pattern that is a pattern of time series data as the input data An input time-series pattern storage network composed of a plurality of nodes having a time-series pattern model, and a plurality of nodes having a time-series pattern model representing a time-series pattern that is a pattern of time-series data as the output data Output time series pattern storage network, and the input time series pattern storage network Of the input time-series pattern storage network in the input / output relation model in which the nodes of the input time-series pattern storage network and the nodes of the output time-series pattern storage network are coupled to each other An input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes, and the input time series pattern storage network is updated in a self-organized manner based on the input winner node. From the recognition learning processing means and the node of the output time series pattern storage network, determine the output winner node that is the most suitable node for the output data of the predetermined time unit, based on the output winner node, Generative learning to update the output time-series pattern storage network in a self-organizing manner A first input winner node for the input data of the predetermined time unit among the input winner nodes of the input time-series pattern storage network and a time of the input data of the predetermined time unit A process of updating the input data of the predetermined time unit delayed by time so as to increase a connection weight indicating a degree of connection with the second input winner node, and the second input of the input time-series pattern storage network A degree of coupling between the winner node and the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit in the output time series pattern storage network And a connection weight updating unit that performs a process of updating the connection weight so as to increase the connection weight.

  A data processing method or program according to a third aspect of the present invention is a data processing method for a data processing apparatus that learns a relationship between input data that is time-series data and output data that is other time-series data, Alternatively, in a program that causes a computer to execute data processing for learning the relationship between input data that is time-series data and output data that is other time-series data, input data in a predetermined time unit is input from the input data. It is composed of a plurality of nodes having a time-series pattern model that extracts the output data of the predetermined time unit from the output data and expresses a time-series pattern that is a pattern of time-series data as the input data. Input time-series pattern storage network and time-series data pattern as the output data An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a pattern, the nodes of the input time-series pattern storage network being coupled to each other, and the input time-series pattern Among the nodes of the input time-series pattern storage network in the input / output relationship model in which the nodes of the storage network and the nodes of the output time-series pattern storage network are combined, the best fit to the input data in the predetermined time unit An input winner node that is a node to be updated, and based on the input winner node, updates the input time-series pattern storage network in a self-organized manner, and selects the predetermined time from the nodes of the output time-series pattern storage network No that best fits the output data of time unit An output winner node is determined, and based on the output winner node, the output time-series pattern storage network is updated in a self-organized manner, and the input time-series pattern storage network of the input winner nodes A first input winner node for input data of a predetermined time unit, and a second input winner node for input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit Update processing to increase the connection weight representing the degree of connection, the second input winner node of the input time series pattern storage network, and the input data of the predetermined time unit of the output time series pattern storage network Of the output winner node for the output data of the predetermined time unit delayed by a certain time from the time of And a process of updating so as to increase the connection weight representing the degree.

  In the third aspect of the present invention as described above, input data in a predetermined time unit is extracted from the input data, and output data in the predetermined time unit is extracted from the output data. Further, an input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the input time series pattern storage network, and the input time series pattern is determined based on the input winner node. The storage network is updated in a self-organized manner, and an output winner node that is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time-series pattern storage network, and the output Based on the winner node, the output time series pattern storage network is updated in a self-organizing manner. And among the input winner nodes of the input time-series pattern storage network, a first input winner node for the input data in the predetermined time unit and a predetermined time from the time of the input data in the predetermined time unit The second input winner node of the input time-series pattern storage network is updated so as to increase the connection weight indicating the degree of connection with the second input winner node for the delayed input data of the predetermined time unit. And a connection weight representing a degree of connection with the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit in the output time series pattern storage network Will be updated to strengthen.

  A data processing device according to a fourth aspect of the present invention is a data processing device that generates output data that is other time-series data with respect to input data that is time-series data. Input data extracting means for extracting input data; an input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data; An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data, and the input time series pattern storage network The nodes are connected to each other, and the input time series pattern storage network Among the nodes of the input time-series pattern storage network in the input / output relationship model in which the nodes of the output time-series pattern storage network are coupled to each other. An input winner node determining means for determining an input winner node that is a node to be determined, and determining a strongest coupled node that is a node having the strongest coupling with the input winner node from the nodes of the input time-series pattern storage network, Generation node determination means for determining, as a generation node for generating the output data, a node having the strongest connection with the strongest connection node among the nodes of the output time series pattern storage network; and the time series possessed by the generation node Generating means for generating the output data based on a pattern model.

  A data processing method or program according to a fourth aspect of the present invention provides a data processing method or a time series for a data processing device that generates output data that is other time-series data for input data that is time-series data. In a program that causes a computer to execute data processing for generating output data that is other time-series data with respect to input data that is data, input data in a predetermined time unit is extracted from the input data, and the input data An input time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a time-series data pattern, and a time-series that is a pattern of time-series data as the output data Output time composed of multiple nodes with a time-series pattern model that represents the pattern A node of the input time series pattern storage network and a node of the output time series pattern storage network are coupled to each other. Determining an input winner node which is a node most suitable for the input data of the predetermined time unit from among the nodes of the input time series pattern storage network in the input / output relation model, Among the nodes, a strongest connection node that is the strongest node with the input winner node is obtained, and a node with the strongest connection with the strongest connection node is selected from the nodes of the output time-series pattern storage network. Determine as a generation node for generating the output data, Based on the time series pattern model consisting node has, comprises generating the output data.

  In the fourth aspect of the present invention as described above, input data in a predetermined time unit is extracted from the input data, and the input data in the predetermined time unit is extracted from the nodes of the input time series pattern storage network. The input winner node that is the node that best fits is determined. Further, a strongest connection node that is a node having the strongest connection with the input winner node is obtained from the nodes of the input time series pattern storage network, and the strongest connection node is obtained from the nodes of the output time series pattern storage network. The node having the strongest connection with the combination node is determined as the generation node that generates the output data. Then, the output data is generated based on the time series pattern model of the generation node.

  According to the first to fourth aspects of the present invention, for example, a robot or the like can perform a task that requires real-time performance.

  Embodiments of the present invention will be described below. Correspondences between the constituent elements of the present invention and the embodiments described in the specification or the drawings are exemplified as follows. This description is intended to confirm that the embodiments supporting the present invention are described in the specification or the drawings. Therefore, even if there is an embodiment which is described in the specification or the drawings but is not described here as an embodiment corresponding to the constituent elements of the present invention, that is not the case. It does not mean that the form does not correspond to the constituent requirements. Conversely, even if an embodiment is described here as corresponding to a configuration requirement, that means that the embodiment does not correspond to a configuration requirement other than the configuration requirement. It's not something to do.

A data processing device according to the first aspect of the present invention provides:
A data processing device that learns the relationship between input data that is time-series data and output data that is other time-series data,
Input / output data extraction means (for example, the data extraction unit 312 in FIG. 32) that extracts the input data in a predetermined time unit from the input data and extracts the output data in the predetermined time unit from the output data;
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage of FIG. 33). Network net ₁ ), and a second input time series pattern storage network (eg, time series pattern storage network net ₃ in FIG. 33);
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 33). ) And
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output Among the nodes of the first input time-series pattern storage network in the input-output relation model (for example, the input-output relation model M _{123 in} FIG. 33) connected to the nodes of the time-series pattern storage network. A first input winner node which is a node most suitable for time unit input data is determined, and the first input time series pattern storage network is updated in a self-organized manner based on the first input winner node. First recognition learning processing means (for example, a recognition learning processing unit 321 in FIG. 32),
A second input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the second input time-series pattern storage network, and based on the second input winner node A second recognition learning processing means for updating the second input time-series pattern storage network in a self-organizing manner (for example, a recognition learning processing unit 421 in FIG. 32);
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. And a generation learning processing means (for example, the generation learning processing unit 322 in FIG.
The first input winner node for the input data of the predetermined time unit of the first input time series pattern storage network and the input data of the predetermined time unit of the second input time series pattern storage network A process of updating so as to increase the coupling weight indicating the degree of coupling with the second input winner node for the input data in the predetermined time unit delayed by a certain time from the time;
The second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the second input time series pattern storage network, and the output time Update the sequence pattern storage network to increase the connection weight indicating the degree of connection with the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit. And a connection weight update unit (for example, a connection weight update unit 433 in FIG. 32).

A data processing method or program according to the first aspect of the present invention includes:
A data processing method for a data processing device that learns the relationship between input data that is time-series data and output data that is other time-series data, or input data that is time-series data and other time-series data Is a program that causes a computer to execute data processing for learning the relationship with output data,
Extracting input data in a predetermined time unit from the input data, and extracting output data in the predetermined time unit from the output data (for example, step S331 in FIG. 35),
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage of FIG. 33). Network net ₁ ), and a second input time series pattern storage network (eg, time series pattern storage network net ₃ in FIG. 33);
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 33). ) And
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output Among the nodes of the first input time-series pattern storage network in the input-output relation model (for example, the input-output relation model M _{123 in} FIG. 33) connected to the nodes of the time-series pattern storage network. A first input winner node which is a node most suitable for time unit input data is determined, and the first input time series pattern storage network is updated in a self-organized manner based on the first input winner node. (For example, step S332 _{1 in} FIG. 35),
A second input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the second input time-series pattern storage network, and based on the second input winner node Updating the second input time-series pattern storage network in a self-organizing manner (for example, step S332 _{2 in} FIG. 35),
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. Is updated in a self-organizing manner (for example, step S332 _{3 in} FIG. 35),
The first input winner node for the input data of the predetermined time unit of the first input time series pattern storage network and the input data of the predetermined time unit of the second input time series pattern storage network A process of updating so as to increase the coupling weight indicating the degree of coupling with the second input winner node for the input data in the predetermined time unit delayed by a certain time from the time;
The second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the second input time series pattern storage network, and the output time Update the sequence pattern storage network to increase the connection weight indicating the degree of connection with the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit. (For example, step S334 in FIG. 35)
Includes steps.

A data processing apparatus according to the second aspect of the present invention provides:
A data processing device that generates output data that is other time-series data for input data that is time-series data,
Input data extraction means (for example, the data extraction unit 315 in FIG. 32) for extracting input data in predetermined time units from the input data;
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage of FIG. 33). Network net ₁ ), and a second input time series pattern storage network (eg, time series pattern storage network net ₃ in FIG. 33);
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 33). ) And
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output Among the nodes of the first input time-series pattern storage network in the input-output relation model (for example, the input-output relation model M _{123 in} FIG. 33) connected to the nodes of the time-series pattern storage network. An input winner node determining means (for example, a winner node determining unit 342 in FIG. 32) for determining an input winner node which is a node most suitable for time-unit input data;
From the nodes of the second input time-series pattern storage network, a strongest connection node that is the strongest node with the input winner node is obtained, and from the nodes of the output time-series pattern storage network, the strongest connection node is obtained. A generation node determination unit (for example, a generation node determination unit 451 in FIG. 32) that determines a node having the strongest connection with a combination node as a generation node that generates the output data;
Generation means (for example, a time series generation unit 361 in FIG. 32) that generates the output data based on the time series pattern model of the generation node.

A data processing method or program according to the second aspect of the present invention includes:
A data processing method of a data processing device that generates output data that is other time series data for input data that is time series data, or other time series data for input data that is time series data A program for causing a computer to execute data processing for generating output data,
Extracting input data in a predetermined time unit from the input data (for example, step S341 in FIG. 36),
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage of FIG. 33). Network net ₁ ), and a second input time series pattern storage network (eg, time series pattern storage network net ₃ in FIG. 33);
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 33). ) And
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output Among the nodes of the first input time-series pattern storage network in the input-output relation model (for example, the input-output relation model M _{123 in} FIG. 33) connected to the nodes of the time-series pattern storage network. An input winner node which is a node most suitable for input data in time units is determined (for example, step S342 in FIG. 36),
From the nodes of the second input time-series pattern storage network, a strongest connection node that is the strongest node with the input winner node is obtained, and from the nodes of the output time-series pattern storage network, the strongest connection node is obtained. A node having the strongest combination with the combination node is determined as a generation node that generates the output data (for example, step S343 in FIG. 36),
The output data is generated based on the time series pattern model of the generation node (for example, step S344 in FIG. 36).
Includes steps.

A data processing device according to the third aspect of the present invention provides:
A data processing device that learns the relationship between input data that is time-series data and output data that is other time-series data,
Input / output data extraction means (for example, the data extraction unit 312 in FIG. 38) for extracting the input data in a predetermined time unit from the input data and extracting the output data in the predetermined time unit from the output data;
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage network net _{1 in} FIG. 37). )When,
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 37) ) And
The input time-series pattern storage network nodes are coupled to each other, and the input time-series pattern storage network node and the output time-series pattern storage network node are coupled to each other. The input winner node which is the node most suitable for the input data of the predetermined time unit is determined from the nodes of the input time series pattern storage network in the input / output relation model M ′ ₁₂₃ ) of FIG. 37, and the input winner is determined. Recognition learning processing means (for example, recognition learning processing unit 321 in FIG. 38) for updating the input time-series pattern storage network in a self-organized manner based on the nodes;
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. A generation learning processing means (for example, a generation learning processing unit 322 in FIG.
Of the input winner nodes of the input time-series pattern storage network, the first input winner node for the input data in the predetermined time unit is delayed by a certain time from the time of the input data in the predetermined time unit Updating to increase the connection weight representing the degree of connection with the second input winner node for the input data in the predetermined time unit;
The second input winner node of the input time series pattern storage network and the output of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the output time series pattern storage network And a connection weight update unit (for example, a connection weight update unit 1433 in FIG. 38) that performs processing to update the connection weight that represents the degree of connection with the output winner node for data.

The data processing method or program according to the third aspect of the present invention provides:
A data processing method for a data processing device that learns the relationship between input data that is time-series data and output data that is other time-series data, or input data that is time-series data and other time-series data Is a program that causes a computer to execute data processing for learning the relationship with output data,
Extracting input data in a predetermined time unit from the input data, and extracting output data in the predetermined time unit from the output data (for example, step S1331 in FIG. 39),
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage network net _{1 in} FIG. 37). )When,
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 37) ) And
The input time-series pattern storage network nodes are coupled to each other, and the input time-series pattern storage network node and the output time-series pattern storage network node are coupled to each other. The input winner node which is the node most suitable for the input data of the predetermined time unit is determined from the nodes of the input time series pattern storage network in the input / output relation model M ′ ₁₂₃ ) of FIG. 37, and the input winner is determined. Based on the node, the input time-series pattern storage network is updated in a self-organizing manner (for example, step S13332 _{1 in} FIG. 39),
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. a self-organizing manner updated (e.g., step S1332 ₂ in FIG. 39),
Of the input winner nodes of the input time-series pattern storage network, the first input winner node for the input data in the predetermined time unit is delayed by a certain time from the time of the input data in the predetermined time unit Updating to increase the connection weight representing the degree of connection with the second input winner node for the input data in the predetermined time unit;
The second input winner node of the input time series pattern storage network and the output of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the output time series pattern storage network And updating so as to increase the connection weight indicating the degree of connection with the output winner node for the data (for example, step S1334 in FIG. 39).
Includes steps.

A data processing device according to the fourth aspect of the present invention provides:
A data processing device that generates output data that is other time-series data for input data that is time-series data,
Input data extraction means (for example, the data extraction unit 315 in FIG. 38) for extracting input data in predetermined time units from the input data;
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage network net _{1 in} FIG. 37). )When,
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 37) ) And
The input time-series pattern storage network nodes are coupled to each other, and the input time-series pattern storage network node and the output time-series pattern storage network node are coupled to each other. 37. Input winner node determination for determining an input winner node which is a node most suitable for the input data in the predetermined time unit from the nodes of the input time series pattern storage network in the input / output relation model M ′ ₁₂₃ ) of FIG. Means (for example, winner node determination unit 342 in FIG. 38);
From the nodes of the input time series pattern storage network, find the strongest coupling node that is the strongest node with the input winner node, and from among the nodes of the output time series pattern storage network, Generation node determination means (for example, the generation node determination unit 1451 in FIG. 38) that determines a node having the strongest combination as a generation node that generates the output data;
Generation means (for example, a time series generation unit 361 in FIG. 38) that generates the output data based on the time series pattern model of the generation node.

According to a fourth aspect of the present invention, there is provided a data processing method or program.
A data processing method of a data processing device that generates output data that is other time series data for input data that is time series data, or other time series data for input data that is time series data A program for causing a computer to execute data processing for generating output data,
Extracting input data in a predetermined time unit from the input data (for example, step S1341 in FIG. 40),
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern that is a pattern of time series data as the input data (for example, the time series pattern storage network net _{1 in} FIG. 37). )When,
An output time series pattern storage network composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as the output data (for example, the time series pattern storage network net _{2 in} FIG. 37) ) And
The input time-series pattern storage network nodes are coupled to each other, and the input time-series pattern storage network node and the output time-series pattern storage network node are coupled to each other. An input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the input time series pattern storage network in the input / output relation model M ′ _{123 of} FIG. 37 (for example, FIG. 40 step S1342),
From the nodes of the input time series pattern storage network, find the strongest coupling node that is the strongest node with the input winner node, and from among the nodes of the output time series pattern storage network, Is determined as a generation node that generates the output data (for example, step S1343 in FIG. 40),
The output data is generated based on the time series pattern model of the generation node (for example, step S1344 in FIG. 40).
Includes steps.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. As a preparation for the previous stage, a time-series pattern storage network will be described.

  FIG. 6 schematically shows an example of a time-series pattern storage network.

  A time-series pattern storage network is a network composed of a plurality of nodes having a time-series pattern model that represents a time-series pattern, and stores the time-series patterns as many as the number of nodes (classifying) in the entire network. To do.

In FIG. 6, the time-series pattern storage network is composed of six nodes N _{1 to} N ₆ .

Each node N _i (in FIG. 6, i = 1, 2,..., 6) constituting the time series pattern storage network has a time series pattern model expressing the time series pattern. Further, the node N _i can have a coupling relationship with other nodes N _j (j = 1, 2,..., 6 in FIG. 6). This connection relationship is called a link. In FIG. 6, for example, the node N ₁ has a direct coupling relationship with the nodes N ₂ and N ₃ . Further, for example, the node N ₃ has a direct coupling relationship with the nodes N ₁ , N ₂ , N ₅ , and N _6, and thus the nodes N ₅ and N ₆ are connected to the node N ₃ via the node N _3. and a N ₁ and indirect coupling relationship. As the binding relationship between the two nodes N _i and N _j, to be considered the shortest binding relationship between the two nodes N _i and N _j.

  Learning of the time-series pattern storage network (learning for storing the time-series pattern in the time-series pattern storage network) is performed using time-series data as learning data for learning. In this respect, the learning of the time-series pattern storage network is greatly different from the supervised learning described with reference to FIG. Moreover, the correct label is not given to the learning data used for learning of the time-series pattern storage network. For this reason, the supervised learning described with reference to FIG. 5 cannot be applied to the learning of the time-series pattern storage network.

  As described above, supervised learning cannot be applied to learning of the time-series pattern storage network, and the type of category and the number of categories of the learning data are unknown. Therefore, learning of the time-series pattern storage network is performed in a self-organized manner so that the features (time-series patterns) of the learning data can be appropriately expressed by the whole (nodes).

  Note that learning of the time-series pattern storage network is unsupervised learning. In addition, learning of the time-series pattern storage network is not always performed so that one certain node corresponds to one certain category. That is, in the time-series pattern storage network, learning may be performed so that one node corresponds to one category, or learning may be performed so that a plurality of nodes correspond to one category. . Further, learning may be performed so that one node corresponds to a plurality of categories. Therefore, even if the learning data cannot be clearly categorized, learning by the time series pattern storage network can be performed.

Next, FIG. 7, when a configuration example of a node N _i of sequence pattern storage network is schematically shown.

The node _Ni includes a time series pattern model 21 that represents a time series pattern and a learning data storage unit 22 that stores learning data used for learning the time series pattern model 21.

Here, in FIG. 7, an HMM (continuous HMM) that is one of the state probability transition models is adopted as the time-series pattern model 21. In FIG. 7, the HMM has left-to-right three states S ₁ , S ₂ , and S ₃ that have only a self-loop and a state transition to the next state (right adjacent state). Yes. In the time-series pattern model 21 in FIG. 7, a circle represents a state, and an arrow represents a state transition. Note that the HMM as the time series pattern model 21 is not limited to the left-to-right type, the three-state type, or the like.

  When the time series pattern model 21 is an HMM as shown in FIG. 7, the HMM as the time series pattern model 21 has a state transition probability and an output probability density function (when the HMM is a discrete HMM, a scalar quantity The probability that a discrete symbol is output).

  The state transition probability is a probability of state transition in the HMM, and is given to each state transition indicated by an arrow in the time series pattern model 21 of FIG. The output probability density function represents the probability density of values observed from the HMM at the time of state transition. For example, a mixed normal distribution is adopted as the output probability density function. These HMM parameters (state transition probability and output probability density function) can be learned (estimated) by, for example, the Baum-Welch method.

At the node _Ni , the statistical characteristics of the learning data stored in the learning data storage unit 22, that is, the time series pattern of the learning data stored in the learning data storage unit 22 is learned in the time series pattern model 21, Thereby, the time-series pattern model 21 and the learning data stored in the learning data storage unit 22 have a correspondence relationship.

Incidentally, when the learning of the sequence pattern storage network, and thus, the time series pattern model 21 learning node N _i is performed by on-line learning for learning each time relative to the time series pattern storage network, the time series data is provided . Thus, the parameters of the time series pattern storage network, that is, if the node N parameters of the time series pattern model 21 of _i (time series pattern model 21 is HMM, as described above, the state transition probability and the output probability density function ) Is updated little by little whenever time-series data is given to the time-series pattern storage network.

  That is, as will be described later, as the learning of the time series pattern storage network proceeds, the learning data stored in the learning data storage unit 22 is updated with the time series data given to the time series pattern storage network. It changes little by little. Then, the learning of the time series pattern model 21 is performed using the learning data that changes little by little, and the parameters of the time series pattern model 21 also change little by little.

  Next, FIG. 8 schematically shows another example of the time-series pattern storage network.

In FIG. 8, the time-series pattern storage network is composed of nine nodes N _{1 to} N ₉ , and these nine nodes N _{1 to} N ₉ are two-dimensionally arranged. That is, in FIG. 8, the nine nodes N _{1 to} N ₉ are arranged on the two-dimensional plane so that the width × length is 3 × 3.

Further, in FIG. 8, there are links (connection relationships) between the nodes adjacent in the horizontal direction and the nodes adjacent in the vertical direction among the nine nodes N _{1 to} N ₉ arranged two-dimensionally. Is given. By providing such a link, it can also be said that the nodes constituting the time-series pattern storage network are given an arrangement structure that is spatially arranged two-dimensionally.

  In a time-series pattern storage network, a distance between two nodes on the space can be defined based on a spatial node arrangement structure given by a link. It can be used as an inter-pattern distance (similarity between time series patterns) of time series patterns expressed by the time series pattern model 21 possessed by each of the two nodes.

  The inter-pattern distance of the time series pattern represented by the distance between the two nodes can be said to be the inter-pattern distance based on the connection relationship (link) between the two nodes.

  As the distance between the two nodes, for example, the number of links constituting the shortest path connecting the two nodes can be employed. In this case, when attention is paid to a certain node, a node having a direct link with the target node (a node adjacent in the horizontal direction or vertical direction of the target node in FIG. 8) is the closest to the target node, A node that can be reached by following a previous link from a node having a direct link with the node becomes farther from the target node as the number of links to be reached increases.

  The links given to the nodes are not limited to those shown in FIGS. The links shown in FIG. 6 and FIG. 8 give a two-dimensional arrangement structure to the node, but the links give other one-dimensional arrangement structures, three-dimensional arrangement structures, and the like. It may be. Furthermore, a link does not necessarily have to be given to a node.

  That is, FIG. 9 schematically shows still another example of the time-series pattern storage network.

In FIG. 9, the time-series pattern storage network is composed of six nodes N _{1 to} N ₆ as in the case of FIG. 6, but all of these six nodes N _{1 to} N ₆ Does not have a link. Therefore, the nodes N _{1 to} N ₆ constituting the time series pattern storage network of FIG. 9 do not have a spatial arrangement structure given by the link. Note that the absence of a spatial arrangement structure means that an arrangement structure without spatial restrictions is given.

  Here, when there is no link between two nodes, the distance between the two nodes in the space cannot be defined, so each of the two nodes (represented by the time series pattern model 21) is expressed. The inter-pattern distance based on the coupling relationship (link) cannot be used as the inter-pattern distance of the time series pattern. Therefore, in this case, for example, a value corresponding to a rank (hereinafter, referred to as “fit rank” as appropriate) based on the degree that a node matches a certain time-series data (its observed value) is used as the inter-pattern distance. it can.

  That is, when certain time-series data is given, the similarity of the time-series pattern expressed by the node to the time-series data can be obtained as the degree of suitability of the node. Now, among the nodes constituting the time-series pattern storage network, for the winner node that is the most suitable node for certain time-series data, each of the winner node and the node with the time-series pattern storage network represents the time series. As the inter-pattern distance of the pattern, a value corresponding to the rank (matching rank) in which the node matches the time series data can be adopted.

  Specifically, among the nodes constituting the time-series pattern storage network, the matching rank of the node that is the winner node is first, and this node (winner node) and the winner node (respectively expressed) The inter-pattern distance (of the time series pattern) can be set to 0 obtained by subtracting 1 from the matching order, for example.

  Further, among the nodes constituting the time-series pattern storage network, the inter-pattern distance between the node having the second highest matching rank and the winner node can be set to 1, for example, by subtracting one from the matching rank. Hereinafter, similarly, a value obtained by subtracting 1 from the matching order of nodes can be set as the inter-pattern distance between the node and the winner node (the node that is the winner).

  Note that the inter-pattern distance represented by the value corresponding to the rank (adaptation rank) based on the degree of conformity with a certain time-series data can be said to be the inter-pattern distance based on the degree that the node conforms to the time-series data. .

  Next, FIG. 10 shows a configuration example of a data processing apparatus that performs various processes using a time-series pattern storage network.

  The signal input unit 1 receives data (hereinafter, referred to as processing target data) that is a target of learning processing and recognition processing described later. Here, the processing target data is, for example, observation values (values (signals) that can be observed from the outside) such as sound, image, LED (Light Emitting Diode) brightness, motor rotation angle and rotation angular velocity. . Further, the processing target data may be, for example, data output from an input device (sensor) that receives an input of a system to which the data processing apparatus of FIG. 10 is applied, or given to an output device that performs some output. May be the data to be generated.

  That is, when the data processing apparatus of FIG. 10 is applied to, for example, a bipedal walking robot and other robots, and the bipedal walking robot performs some processing according to an external situation, the signal input unit 1 is It can be composed of sensors that sense external conditions. Specifically, the signal input unit 1 can be composed of, for example, a microphone (microphone) or a camera.

  When the signal input unit 1 is configured by a microphone, externally generated voice (including human voice, animal squeal, object sound, and all other sounds) is biped with respect to the microphone. Input as input data to the robot (system to which the data processing apparatus is applied), and corresponding audio data is supplied to the feature extraction unit 2. When the signal input unit 1 is configured by a camera, external light is input to the camera as input data for the biped robot, and corresponding image data is input to the feature extraction unit 2. Supplied.

  In addition, when the biped walking robot can move, for example, a part corresponding to a hand or a leg by a motor as an actuator, the signal input unit 1 measures the rotation angle and the rotation speed of the motor. An apparatus (a sensor for sensing a rotation angle and a rotation speed) can be used. The motor that moves the part corresponding to the hand or foot of the biped robot gives a driving signal as an electric signal for rotationally driving the motor and moves the part corresponding to the hand or foot by applying force from the outside. However, the measuring device can measure even a rotation angle or a rotation speed caused by any rotation.

  When the signal input unit 1 is configured by a measuring device, a signal representing the rotation angle and rotation speed of the motor is input to the measuring device as output data from the biped walking robot and measured. The measurement result is supplied to the feature extraction unit 2.

  The processing target data input to the signal input unit 1 may be stationary data (stationary signal) with a constant temporal change, or non-stationary data (non-steady with a temporal change not constant). (Stationary signal).

  In the following description, it is assumed that, for example, voice that is one of time-series data is input to the signal input unit 1. Furthermore, it is assumed that only the audio data of the so-called audio section is supplied from the signal input unit 1 to the feature extraction unit 2. Note that the method for detecting the voice section is not particularly limited. Also, the audio data supplied from the signal input unit 1 to the feature extraction unit 2 does not necessarily have to be the length of the audio section, and may be divided into an appropriate length. That is, the speech data supplied from the signal input unit 1 to the feature extraction unit 2 may be, for example, a phoneme or phoneme unit, or a word or sentence, or from one punctuation mark to the next punctuation mark. good.

  Here, the processing target data supplied from the signal input unit 1 to the feature extraction unit 2 is not limited to audio data, and the section is not particularly limited. That is, it is only necessary that the signal input unit 1 supplies the feature extraction unit 2 with the processing target data divided into appropriate lengths by the best method. Note that the data to be processed (section) supplied from the signal input unit 1 to the feature extraction unit 2 may or may not be constant.

  The feature extraction unit 2 extracts a feature amount from audio data that is time-series data as processing target data from the signal input unit 1, and recognizes a time-series feature amount that is time-series data obtained as a result of the recognition unit 3. And supplied to the learning unit 4. That is, the feature extraction unit 2 performs processing such as frequency analysis on the audio data from the signal input unit 1 at regular time intervals, and, for example, features such as a mel cepstrum coefficient (MFCC (Mel Frequency Cepstrum Coefficient)). The time series data of the mel cepstrum coefficients is extracted and supplied to the recognition unit 3 and the learning unit 4. Note that the time series data supplied from the feature extraction unit 2 to the recognition unit 3 and the learning unit 4 is also an observation value that can be observed from the outside.

  The recognition unit 3 recognizes the time series data supplied from the feature extraction unit 2 based on the time series pattern storage network stored in the storage unit 5 and outputs the recognition result.

  Here, the learning unit 4 updates the time-series pattern storage network stored in the storage unit 5 in a self-organized manner based on the time-series data (observed values) supplied from the feature extraction unit 2. That is, the learning unit 4 updates the parameters of the time series pattern storage network stored in the storage unit 5 based on the time series data supplied from the feature extraction unit 2. Note that this parameter update may be referred to as learning.

  When the learning unit 4 repeatedly gives time-series data to which no correct answer label is assigned, there is no teacher to acquire a characteristic pattern (time-series pattern) in the given time-series data in a self-organizing manner. Learning is performed. As a result, a representative time series pattern is efficiently stored in the time series pattern storage network of the storage unit 5. That is, the time series data supplied from the feature extraction unit 2 to the recognition unit 3 and the learning unit 4 can be classified into several patterns (time series patterns). In the learning unit 4, the time series pattern storage network includes Learning for storing a representative time series pattern of time series data is performed.

  The storage unit 5 stores a time-series pattern storage network, and the time-series pattern storage network (its parameters) is appropriately updated by the learning unit 4.

  Control data is supplied to the generation unit 6. The control data supplied to the generation unit 6 represents any one of the time series patterns stored in the time series pattern storage network of the storage unit 5 (a node label described later). Based on the time series pattern storage network of the unit 5, time series data of the time series pattern represented by the control data supplied thereto is generated and output.

  Next, FIG. 11 shows a configuration example of the learning unit 4 of FIG.

  The learning unit 4 includes a time series data storage unit 31 and a learning processing unit 32.

  The time series data storage unit 31 is supplied with a series of feature amounts as new time series data (one section) from the feature extraction unit 2. The time series data storage unit 31 temporarily stores new time series data from the feature extraction unit 2 (until processing using the new time series data by the learning processing unit 32 is completed).

  The learning processing unit 32 updates the time-series pattern storage network stored in the storage unit 5 in a self-organized manner based on the new time-series data (observed values) stored in the time-series data storage unit 31. .

  Next, FIG. 12 shows a configuration example of the learning processing unit 32 of FIG.

  The score calculation unit 41 is adapted to the time-series data (observed values) stored in the time-series data storage unit 31 for each node constituting the time-series pattern storage network stored in the storage unit 5. The degree is obtained as a score and supplied to the winner node determination unit 42. That is, when the time series pattern model 21 possessed by the node is an HMM as shown in FIG. 7, for example, the score calculation unit 41 obtains time series data from the HMM as the time series pattern model 21 possessed by the node. The likelihood that the time-series data stored in the storage unit 31 is observed is obtained and supplied to the winner node determination unit 42 as the score of the node.

  The winner node determination unit 42 obtains a node that best matches the time-series data stored in the time-series data storage unit 31 in the time-series pattern storage network stored in the storage unit 5, and determines that node as a winner node. To do.

  That is, the winner node determination unit 42 determines the node having the highest score from the score calculation unit 41 among the nodes constituting the time-series pattern storage network stored in the storage unit 5 as the winner node. Then, the winner node determination unit 42 supplies information representing the winner node to the weight determination unit 43.

  Here, a node label, which is a label for identifying each node, can be attached to the nodes constituting the time-series pattern storage network. And a node label is employable as information showing a winner node and information showing other nodes. The node label is a label for identifying the node itself, and has nothing to do with the correct answer label indicating what the correct answer is.

  The weight determination unit 43 determines an update weight to be described later for each node constituting the time-series pattern storage network stored in the storage unit 5 based on the winner node represented by the node label supplied from the winner node determination unit 42. And supplied to the learning data updating unit 44.

  That is, the weight determination unit 43 calculates the update weight of each node (including the winner node) included in the time-series pattern storage network stored in the storage unit 5 based on the inter-pattern distance between the node and the winner node. It is determined and supplied to the learning data update unit 44.

  Here, the time-series pattern model 21 (FIG. 7) of the node is updated using new time-series data stored in the time-series data storage unit 31 (FIG. 11). The degree of the influence of the new time series data that the time series pattern model 21 receives by updating the time series pattern model 21 of the node. Therefore, if the update weight of a node is 0, the time series pattern model 21 possessed by the node is not affected (not updated) by new time series data.

  As the inter-pattern distance when the weight determining unit 43 determines the update weight of each node constituting the time-series pattern storage network stored in the storage unit 5, the nodes of the time-series pattern storage network are shown in FIG. As shown in FIG. 8 and FIG. 8, when there is a link, the inter-pattern distance based on the connection relationship between the node and the winner node is adopted, and the nodes of the time-series pattern storage network are shown in FIG. Thus, when there is no link, the inter-pattern distance based on the degree to which the node matches the new time-series data stored in the time-series data storage unit 31 (FIG. 11) can be adopted.

  That is, the weight determination unit 43 refers to the time-series pattern storage network stored in the storage unit 5 and combines each node of the time-series pattern storage network with the winner node represented by the node label from the winner node determination unit 42. An inter-pattern distance based on the relationship is obtained, and an update weight for each node of the time-series pattern storage network is determined based on the inter-pattern distance.

  Alternatively, the weight determination unit 43 refers to the time series pattern storage network stored in the storage unit 5, and for each node of the time series pattern storage network, a new time stored in the time series data storage unit 31. For example, a score similar to that obtained by the score calculation unit 41 is obtained as the degree of matching with the series data. Furthermore, the weight determination unit 43 obtains a value corresponding to the rank (matching rank) based on the score of each node of the time-series pattern storage network as an inter-pattern distance based on the degree of matching with the new time-series data. Based on the distance between the patterns, the update weight of each node of the time-series pattern storage network is determined.

  The score of the node may be obtained by the weight determining unit 43, but may be supplied from the score calculating unit 41 to the weight determining unit 43.

  The learning data update unit 44 updates the learning data stored in the learning data storage unit 22 (FIG. 7) included in each node of the time-series pattern storage network stored in the storage unit 5.

  That is, the learning data update unit 44 uses the learning data already stored in the learning data storage unit 22 included in the node and the time series data stored in the time series data storage unit 31 from the weight determination unit 43. The mixing is performed according to the update weight of the corresponding node, and the mixing result is stored in the learning data storage unit 22 as new learning data, thereby updating the storage content of the learning data storage unit 22.

  As described above, when the learning data stored in the learning data storage unit 22 (FIG. 7) is updated according to the update weight, the learning data update unit 44 notifies the model learning unit that the update has been completed. 45.

  When the model learning unit 45 receives an end notification from the learning data update unit 44, the model learning unit 45 uses the learning data stored in the learning data storage unit 22 (FIG. 7) updated by the learning data update unit 44 to use the time series pattern. By learning the time series pattern model 21 of each node of the storage network, the time series pattern model 21 is updated.

  Therefore, the model learning unit 45 updates the time-series pattern model 21 possessed by a node with (part of) the learning data stored in the learning data storage unit 22 (FIG. 7) possessed by the node and the time-series data. This is performed based on new time series data stored in the storage unit 31. In addition, since the storage content of the learning data storage unit 22 is updated according to the update weight, it can be said that the update of the time series pattern model 21 by the model learning unit 45 is performed based on the update weight.

  Next, FIG. 13 shows a determination method for determining the update weight in the weight determination unit 43 of FIG.

  The weight determination unit 43, for example, according to a curve (hereinafter referred to as a distance / weight curve) representing a relationship in which the update weight α decreases with respect to an increase in the inter-pattern distance d with the winner node as shown in FIG. An update weight (node update weight) α for the node is determined. According to the distance / weight curve, a node with a shorter inter-pattern distance d with the winner node determines a larger update weight α, and a node with a longer inter-pattern distance d determines a smaller update weight α.

  In the distance / weight curve of FIG. 13, the horizontal axis (from left to right) represents the update weight α, and the vertical axis (from top to bottom) represents the inter-pattern distance d.

In FIG. 13, as the inter-pattern distance d, for example, the inter-pattern distance based on the connection relationship with the node, that is, the distance from the winner node is adopted, and the time-series pattern storage network is configured along the vertical axis. Six nodes N _{1 to} N ₆ are described at positions (positions on the vertical axis) corresponding to the distance between each node _Ni and the winner node.

In FIG. 13, the six nodes N _{1 to} N ₆ constituting the time-series pattern storage network are closer to the winner node in that order. Among the _six nodes N _{1 to} N ₆ constituting the time-series pattern storage network, the node closest to the winner node, that is, the node N ₁ having a distance of 0 from the winner node is the winner node ( Node).

Here, when the time-series pattern storage network has a two-dimensional arrangement structure as shown in FIG. 8 and the winner node is, for example, the node N ₆ , the winner node N ₆ The distance to the node N ₆ is 0 which is the closest (first), and the inter-pattern distance d between the node N ₆ and the winner node N ₆ is also 0. The pattern of the winning node N _6, node N _3, N ₅ or N ₉ distances respectively, is 1 second closest, a node N _3, N ₅ or N ₉ respectively winning node N _6, The distance d is also 1. Moreover, the pattern of the winning node N _6, node N _2, N ₄ or N ₈ distance respectively, is 2 close to the third, and the node N _2, N ₄ or N ₈ respectively winning node N _6, The distance d is also 2. Further, the winning node N _6, node N ₁ or N ₇ the distance between each is farthest (close to 4 th) 3, also the inter-pattern distance d between the respective nodes N ₁ or N ₇ and winning node N ₆ 3

  On the other hand, for the time-series pattern storage network shown in FIG. 9 where the node does not have a link, for example, the inter-pattern distance based on the degree to which the node matches the new time-series data, that is, the time-series data with the new A value corresponding to the rank (matching rank) based on the degree of matching to is obtained as the inter-pattern distance d between the node and the winner node. That is, in this case, the inter-pattern distance d between the node (winner node) having the highest score (the highest) and the winner node is 0, and the inter-pattern distance d between the node having the second highest score and the winner node Is set to 1. Hereinafter, similarly, the inter-pattern distance d between the node having the highest score and the winner node is k−1.

  Next, for example, a distance / weight curve as shown in FIG. 13 representing the relationship between the update weight α and the inter-pattern distance d is given by, for example, Expression (3).

  Here, in Expression (3), the constant G is a constant representing the update weight of the winner node, and γ is an attenuation coefficient, and is a constant in the range of 0 <γ <1. Further, the variable Δ is an update weight α of a node in the vicinity of the winner node (a node having a short inter-pattern distance d to the winner node) when the above-described SMA is adopted as an update method for updating the time-series pattern storage network. It is a variable for adjusting.

  As described above, 0 is given as the inter-pattern distance d of the node that is the winner node, and hereinafter, as the inter-pattern distance d of other nodes, according to the distance to the winner node or the matching order, 2... Is given, in Equation (3), for example, if G = 8, γ = 0.5, and Δ = 1, the update weight α of the node that is the winner node is 8 (= G ) Is required. Hereinafter, as the distance to the winner node or the matching order increases, the update weight α of the node is determined to be a value that decreases as 4, 2, 1,.

  Here, when the attenuation coefficient Δ in Expression (3) is a large value, the change of the update weight α with respect to the change of the inter-pattern distance d becomes moderate, and conversely, the attenuation coefficient Δ is a value close to 0. In this case, the change in the update weight α with respect to the change in the inter-pattern distance d becomes steep.

  Accordingly, when the attenuation coefficient Δ is adjusted to gradually approach 0 from 1 as described above, for example, the change in the update weight α with respect to the change in the inter-pattern distance d becomes steep, and the update weight α is As the distance d increases, the value decreases. When the attenuation coefficient Δ is close to 0, the update weight α of the nodes other than the winner node (the node that is the winner node) is almost 0. In this case, the update method for updating the time-series pattern storage network is described above. This is (almost) equivalent to using WTA.

  In this way, by adjusting the attenuation coefficient Δ, it is possible to adjust the update weight α of the node in the vicinity of the winner node when SMA is adopted as the update method for updating the time-series pattern storage network.

  For example, the attenuation coefficient Δ may be set to a large value at the start of updating (learning) of the time-series pattern storage network, and may be set to a small value as time passes, that is, as the number of updates increases. it can. In this case, at the start of updating the time-series pattern storage network, the update weight α of each node of the time-series pattern storage network is determined according to a distance / weight curve in which the change of the update weight α with respect to the change in the inter-pattern distance d is gentle. As the update (learning) progresses (proceeds), the update of each node of the time-series pattern storage network is performed according to the distance / weight curve in which the change of the update weight α with respect to the change of the inter-pattern distance d becomes steep. A weight α is determined.

  That is, in this case, the update of the time series pattern model 21 of the winner node is performed regardless of the progress of learning (updating) of the new time series data stored in the time series data storage unit 31 (FIG. 12). It is done to be strongly influenced. On the other hand, the update of the nodes other than the winner node (the time series pattern model 21) has a relatively large range of nodes (a node having a small inter-pattern distance d with the winner node) at the start of learning. (Up to the node) so as to be influenced by the new time-series data. Then, as learning progresses, the update of nodes other than the winner node is performed so as to be influenced by new time-series data only for nodes in a narrow range gradually.

  The weight determination unit 43 in FIG. 12 determines the update weight α of each node of the time-series pattern storage network as described above, and the learning data update unit 44 stores it in the learning data storage unit 22 of each node. Learning data is updated based on the update weight α of the node.

  Next, an update method for updating learning data stored in the learning data storage unit 22 included in the node will be described with reference to FIG.

Now, the learning data storage unit 22 included in a node N _i, learning data is already stored, the time series pattern model 21 of the node N _i is a learning data in the learning data storage unit 22 has already stored It is assumed that learning has been performed using this.

Learning data updating unit 44, as described above, the node N _i already the stored learning data in the learning data storage unit 22 included in the (hereinafter referred to as old learning data) and the time series data storage unit 31 (FIG. the time series data stored new 12), were mixed in accordance with the updating weight α of the node N _i from the weight determiner 43, the mixing result as a new learning data, stored in the learning data storage unit 22 By doing so, the storage content of the learning data storage unit 22 is updated to new learning data.

In other words, the learning data update unit 44 adds new time series data to the old learning data to obtain new learning data in which the old learning data and the new time series data are mixed. additional time series data new to the data (mixed with the old training data and the new time series data) is performed according to the ratio corresponding to the updating weight α of the node N _i.

Here, the node updates the time-series pattern model 21 (FIG. 7) of the N _i are to be done by a learning using a new learning data, varying the ratio of mixing the new time series data and the old learning data Thus, the degree of influence (intensity) of the new time series data received by the time series pattern model 21 by the update can be changed.

In the node N _i, as a percentage of mixing the new time series data and the old learning data, a node value corresponding to the updating weight α of N _i is employed, for example, as the updating weight α is large, the new The value is such that the ratio of the time series data is large (the ratio of the old learning data is small).

Specifically, in the learning data storage unit 22 of the node N _i is intended to time-series data of a certain number (learning data) is stored, the number of its constant and H. In this case, the learning of the time series pattern model 21 of the node N _i is always carried out using the H-number of learning data (time-series data).

When a constant number H of learning data is always stored in the learning data storage unit 22, the number of new learning data obtained by mixing new time-series data and old learning data needs to be H. There is such a mixture of the new time series data and the old learning data, as a method of performing in a ratio corresponding to the updating weight α of the node N _i, and a new time series data and the old learning data, There is a method of mixing at a ratio α: H-α.

  As a specific method for mixing new time-series data and old learning data at a ratio α: H-α, as shown in FIG. 14, H-α old out of H old learning data are used. There is a method of obtaining H new learning data by adding α new time-series data to the learning data.

In this case, the number H of the time series data stored in the learning data storage unit 22, for example, a 100, node N _i updating weight α of, for example, if 8, the stored contents of the learning data storage unit 22 Is updated to 100 new learning data obtained by adding 8 new time-series data to 92 old learning data out of 100 old learning data.

  There is a method of adding α new time series data to H-α old learning data out of H old learning data after waiting for α new time series data to be obtained. However, with this method, the stored content of the learning data storage unit 22 cannot be updated every time one new time series data is obtained.

  Therefore, the update of the stored contents of the learning data storage unit 22 adds only α new time series data to the H-α old learning data every time one new time series data is obtained. Can be done. That is, one new time-series data is copied to α new time-series data, and the α new time-series data is excluded from the H old learning data in the order of oldness. By adding to the remaining H-α old learning data, the storage content of the learning data storage unit 22 is updated. Thereby, whenever one new time series data is obtained, the memory content of the learning data memory | storage part 22 can be updated.

  As described above, by updating the storage contents of the learning data storage unit 22, only the H time-series data in the new order are always held as learning data in the learning data storage unit 22. Thus, the ratio (ratio) of new time-series data in the learning data is adjusted by the update weight α.

  Next, a learning process for learning the time-series pattern storage network performed by the data processing apparatus of FIG. 10 will be described with reference to the flowchart of FIG.

  First, in step S1, the learning processing unit 32 of the learning unit 4 (FIG. 11) has parameters of the time series pattern storage network stored in the storage unit 5, that is, each node of the time series pattern storage network has. For example, initialization processing for initializing parameters of the HMM as the sequence pattern model 21 (FIG. 7) is performed. By this initialization process, appropriate initial values are given as parameters (state transition probability and output probability density function) of the HMM. In the initialization process, how the initial values are given in the HMM parameters is not particularly limited.

  Thereafter, in step S2, when one piece of processing target data, that is, for example, voice data of one speech section is input to the signal input unit 1, the signal input unit 1 extracts the processing target data from the feature extraction unit. Supply to part 2. The feature extraction unit 2 extracts feature amounts from the processing target data, and supplies time series data (one new time series data) of the feature amounts to the learning unit 4.

  The learning unit 4 (FIG. 11) temporarily stores new time-series data from the feature extraction unit 2 in the time-series data storage unit 31, and hereinafter, the time series stored in the storage unit 5 in steps S3 to S7. The pattern storage network is updated (learned) in a self-organized manner based on the new time-series data (observed values) stored in the time-series data storage unit 31.

  That is, in the learning processing unit 32 (FIG. 12) of the learning unit 4, in step S 3, the score calculation unit 41 reads out new time series data stored in the time series data storage unit 31 and stores it in the storage unit 5. For each node constituting the time-series pattern storage network, a score representing the degree to which the node matches the new time-series data is obtained.

  Specifically, when the time series pattern model 21 (FIG. 7) possessed by the node is, for example, an HMM, the log likelihood that new time series data is observed is obtained from the HMM as a score. Here, as a log likelihood calculation method, for example, a Viterbi algorithm can be employed.

  If the score calculation part 41 calculates the score with respect to new time series data about all the nodes which a time series pattern storage network has, the score about each node will be supplied to the winner node determination part 42. FIG.

  In step S4, the winner node determination unit 42 obtains the node having the highest score from the score calculation unit 41 among the nodes constituting the time-series pattern storage network, and determines that node as the winner node. Then, the winner node determination unit 42 supplies a node label as information representing the winner node to the weight determination unit 43.

  In step S5, the weight determination unit 43 determines the update weight of each node constituting the time-series pattern storage network with the winner node represented by the node label from the winner node determination unit 42 as a reference.

  That is, as described with reference to FIG. 13, the weight determination unit 43 determines that the change in the update weight α with respect to the change in the inter-pattern distance d becomes steeper as the update (learning) of the time-series pattern storage network proceeds. In accordance with the distance / weight curve represented by (3), the update weight α of each node of the time-series pattern storage network is determined and supplied to the learning data update unit 44.

  In step S6, the learning data update unit 44 updates the learning data stored in the learning data storage unit 22 (FIG. 7) of each node of the time-series pattern storage network from the weight determination unit 43. Update according to weight. That is, the learning data update unit 44, as described in FIG. 14, includes new time series data stored in the time series data storage unit 31 and old learning data stored in the node learning data storage unit 22. Are mixed at a ratio α: H−α corresponding to the update weight α of the node to obtain H new learning data, and the learning data storage unit 22 uses the H new learning data. Update the stored contents of.

  When the learning data update unit 44 updates the stored contents of the learning data storage unit 22 (FIG. 7) of all the nodes of the time-series pattern storage network, the learning data update unit 44 supplies an end notification to the model learning unit 45 indicating that the update has been completed. .

  When the model learning unit 45 receives an end notification from the learning data update unit 44, the model learning unit 45 updates the parameters of the time-series pattern storage network in step S7.

  That is, the model learning unit 45 uses the new learning data stored in the learning data storage unit 22 after being updated by the learning data update unit 44 for each node of the time series pattern storage network. By performing learning, the time series pattern model 21 is updated.

  Specifically, when the time-series pattern model 21 possessed by a node is, for example, an HMM, HMM learning is performed using new learning data stored in the learning data storage unit 22 possessed by the node. . In this learning, for example, the current state transition probability and output probability density function of the HMM are used as initial values, and new state transition probabilities and output probability density functions are obtained by the Baum-Welch method using new learning data. It is done. Then, the state transition probability and output probability density function of the HMM are updated by the new state transition probability and output probability density function, respectively.

  Thereafter, the process returns from step S7 to step S2, waits for the next data to be processed to be input to the signal input unit 1, and thereafter the same processing is repeated.

  According to the learning process of FIG. 15, when one new time series data is obtained, a winner node is determined from the nodes constituting the time series pattern storage network for the new time series data. (Step S4). Further, the update weight of each node constituting the time-series pattern storage network is determined based on the winner node (step S5). Based on the update weight, the parameters of the time series pattern model 21 (FIG. 7) possessed by each node constituting the time series pattern storage network are updated.

  That is, in the learning process of FIG. 15, the parameters of the nodes constituting the time-series pattern storage network are updated once for one new time-series data. Is obtained by repeating updating of the parameter of the node each time.

  When the learning is sufficiently performed, the time series pattern model 21 included in each node of the time series pattern storage network learns (acquires) a certain time series pattern. Since the number (type) of time series patterns learned in the entire time series pattern storage network matches the number of nodes included in the time series pattern storage network, the number of nodes included in the time series pattern storage network is, for example, 100. In some cases, 100 types of time-series patterns are learned. Based on this time series pattern, the recognition unit 3 (FIG. 10) can perform recognition processing for recognizing time series data (processing target data), and the generation unit 6 (FIG. 10) generates time series data. It is possible to perform the generation process.

  Next, FIG. 16 shows a configuration example of the recognition unit 3 of FIG.

  As described with reference to FIG. 10, (one piece) of time series data is supplied to the recognition unit 3 from the feature extraction unit 2, and this time series data is supplied to the score calculation unit 51.

  Similar to the score calculation unit 41 of the learning processing unit 32 (FIG. 12), the score calculation unit 51 is configured such that each node constituting the time-series pattern storage network stored in the storage unit 5 is the feature extraction unit 2. A score representing the degree of conformity to the time-series data (observed values) from is obtained and supplied to the winner node determination unit 52.

  Similarly to the winner node determination unit 42 of the learning processing unit 32 (FIG. 12), the winner node determination unit 52 uses the time series data from the feature extraction unit 2 in the time series pattern storage network stored in the storage unit 5. Find a matching node and determine that node as the winner node.

  That is, the winner node determination unit 52 determines the node having the highest score from the score calculation unit 51 among the nodes constituting the time-series pattern storage network stored in the storage unit 5 as the winner node. Then, the winner node determination unit 52 supplies the output unit 53 with a node label as information representing the winner node.

  The output unit 53 inputs the node label representing the winner node from the winner node determination unit 52 to the signal input unit 1 corresponding to the time-series data of the feature amount from the feature extraction unit 2 and, consequently, the feature amount. Output as recognition result of data to be processed.

  Note that the score calculation unit 51 of the recognition unit 3 and the score calculation unit 41 of the learning processing unit 32 (FIG. 12) can be shared by either one of the score calculation units. The same applies to the winner node determination unit 52 of the recognition unit 3 and the winner node determination unit 42 of the learning processing unit 32 (FIG. 12).

  Next, a recognition process for recognizing time-series data performed by the data processing apparatus of FIG. 10 will be described with reference to the flowchart of FIG.

  In step S21, when one piece of processing target data, that is, for example, voice data (time-series data) of one voice section is input to the signal input unit 1, the signal input unit 1 selects the processing target data. To the feature extraction unit 2. The feature extraction unit 2 extracts feature amounts from the time series data that is the processing target data, and supplies the time series data of the feature amounts to the recognition unit 3.

  In the recognizing unit 3 (FIG. 16), in step S22, the score calculating unit 51 matches each node constituting the time-series pattern storage network stored in the storage unit 5 with the time-series data from the feature extracting unit 2. A score representing the degree is obtained and supplied to the winner node determination unit 52.

  In step S23, the winner node determination unit 52 obtains the node having the highest score from the score calculation unit 51 among the nodes constituting the time-series pattern storage network, and determines that node as the winner node. Then, the winner node determination unit 52 supplies the output unit 53 with a node label as information representing the winner node.

  The output unit 53 outputs the node label from the winner node determination unit 52 as the recognition result of the time-series data (processing target data input to the signal input unit 1) from the feature extraction unit 2, and ends the processing.

  Note that the node label output from the output unit 53 (the node label of the winner node) can be supplied to the generating unit 6 as control data, for example.

  According to the recognition processing using the time series pattern storage network as described above, a fine recognition result according to the number of nodes of the time series pattern storage network can be obtained.

  That is, for example, it is assumed that learning of the time-series pattern storage network is performed using voice data obtained by utterances of three categories “A”, “B”, and “C”.

  When the audio data of the three categories “A”, “B”, and “C” used for learning of the time-series pattern storage network includes the utterances of a large number of speakers, for example, the audio of the category “A” Even for data, there are various variations of voice data depending on differences in speech rate, intonation, speaker age, gender, and the like.

  For supervised learning, learning using audio data of categories “A”, “B”, and “C” is performed only for audio data of category “A”, only for audio data of category “B”, and category “C”. This is performed using only the audio data. Therefore, the learning result of each category cannot be varied due to the difference in the speech speed or the like.

  On the other hand, in learning of a time-series pattern storage network, audio data of categories “A”, “B”, and “C” are used without distinguishing (classifying) them. Then, in the time series pattern storage network, as described above, time series patterns as many as the number of nodes of the time series pattern storage network are learned.

  Accordingly, in the time-series pattern storage network, for example, if the number of nodes included in the time series pattern storage network is greater than 3, even if the audio data of one category “A” is included in the audio data of the category “A” One variation (a time series pattern) may be learned at one node, and the other variation may be learned at another node.

  As described above, when various variations of the audio data of the category “A” are learned in the plurality of nodes, in the recognition process, for example, the audio data of the category “A” is input as the processing target data. Among the plurality of nodes from which the speech data of category “A” has been learned, the node that best matches the processing target data is determined as the winner node, and the node label of the winner node is output as the recognition result.

  That is, in the recognition process using the time-series pattern storage network, it is not determined which category of audio data is the category “A”, “B”, or “C”. Then, it is determined which of the time-series patterns corresponding to the number of the nodes acquired by learning the time-series pattern storage network is most suitable (similar).

  That is, in learning of the time-series pattern storage network, a time-series pattern is acquired with a fineness according to the number of nodes of the time-series pattern storage network, and in the recognition using the time-series pattern storage network, the time-series pattern storage network The time series data is classified (classified) with fineness according to the number of nodes included in the.

  For each node in the time-series pattern storage network where (sufficient) learning has been performed, the category “A”, “B”, “C” is appropriately selected according to the time-series pattern acquired by the node. If a correct answer label is given, in the recognition process using the time-series pattern storage network, the time-series data (data to be processed) is audio data of any of the categories “A”, “B”, and “C”. The recognition result can be obtained.

  Next, FIG. 18 illustrates a configuration example of the generation unit 6 of FIG.

  As described with reference to FIG. 10, control data is supplied to the generation unit 6. The control data supplied to the generation unit 6 represents one of the time series patterns stored in the time series pattern storage network of the storage unit 5 and thus the nodes constituting the time series pattern storage network. For example, a node label.

  The control data supplied to the generation unit 6 is supplied to the generation node determination unit 61. The generation node determination unit 61 uses, in the time-series pattern storage network stored in the storage unit 5, a node represented by control data supplied thereto to generate time-series data (hereinafter referred to as a generation node as appropriate). And the determination result is supplied to the time-series generation unit 62.

  That is, the generation node determination unit 61 determines the node represented by the node label as the control data supplied thereto as a generation node, and supplies the determination result to the time series generation unit 62.

  The time series generation unit 62 generates time series data based on the determination result from the generation node determination unit 61 based on the time series pattern model 21 (FIG. 7) of the generation node and supplies the time series data to the output unit 63.

  Here, when the time series pattern model 21 is, for example, an HMM, the time series generation unit 62 represents the likelihood that the time series data is observed in the HMM as the time series pattern model 21 of the generation node. Generate time-series data that maximizes output probability. As for the generation of time-series data using the HMM, for example, there is a method of generating time-series data that smoothly changes by using a dynamic feature amount, and the time-series generation unit 62 uses the method, Time series data can be generated. For example, K. Tokuda, T. Yoshimura, T. Masuko, T. Kobayashi, T. Kitamura, "SPEECH PARAMETER GENERATION ALGORITHMS FOR HMM-BASED SPEECH SYNTHESIS", Proc. Of ICASSP 2000, vol.3, pp.1315-1318, June 2000.

  As another method for generating time series data using an HMM, for example, time series data generation based on probabilistic trials using parameters of the HMM is repeatedly performed, and the average is taken to generate time series data from the HMM. Methods for generating serial data have also been proposed. For details, see, for example, Tetsuya Inagi, Hiroaki Tanie, Yoshihiko Nakamura, “Keyframe Extraction and Reconstruction of Time Series Data Using Continuously Distributed Hidden Markov Models” ”, Proceedings of the Japan Society of Mechanical Engineers Robotics and Mechatronics Lecture 2003, 2P1-3F-C6, 2003.

  The output unit 63 converts the time series data from the time series generation unit 62 into time series data corresponding to the processing target data and outputs the time series data. That is, the time series data generated by the time series generation unit 62 is time series data of feature amounts used for learning the time series pattern model 21 possessed by the node, and the output unit 63 outputs time series of the feature amounts. Data is converted into data to be processed (corresponding data) and output.

  Specifically, for example, if the processing target data is voice data and the feature extraction unit 2 extracts a mel cepstrum coefficient from the voice data as a feature quantity, the time series data of the mel cepstrum is a time series pattern model. Therefore, the time series data generated by the time series generation unit 62 based on the time series pattern model 21 (FIG. 7) of the generation node is the time series data of the mel cepstrum. The output unit 63 converts the mel cepstrum (time series data) generated by the time series generation unit 62 into audio data that is time series data corresponding to the processing target data.

  As a method of converting time series data of mel cepstrum into voice data (time domain voice), for example, the mel cepstrum time series data is filtered by a synthesis filter called MLSA filter (Mel logarithm spectrum approximation filter). There is a way. For details on MLSA filters, see, for example, Kiyoshi Imai, Kazuo Sumita, Chieko Furuichi, “Mel Log Spectrum Approximation (MLSA) Filter for Speech Synthesis”, IEICE Transactions (A), J66-A, 2, pp.122-129, 1983, Keiichi Tokuda, Takao Kobayashi, Hironori Saito, Toshiaki Fukada, Kiyoshi Imai, "Spectrum estimation of speech using mel cepstrum as a parameter", IEICE Transactions (A), J74-A , 8, pp.1240-1248, 1991.

  Next, generation processing for generating time-series data (processing target data) performed by the data processing apparatus of FIG. 10 will be described with reference to the flowchart of FIG.

  In step S31, the control data is input to the generation unit 6 (FIG. 18). This control data is supplied to the generation node determination unit 61. The generation node determination unit 61 determines the node represented by the node label as the control data among the nodes constituting the time-series pattern storage network stored in the storage unit 5 as the generation node, and the determination result is This is supplied to the sequence generation unit 62.

  In step S33, the time series generation unit 62 follows the determination result from the generation node determination unit 61, and the time series pattern model that the generation node of the nodes constituting the time series pattern storage network stored in the storage unit 5 has. 21 (time parameter) is generated and supplied to the output unit 63. In step S <b> 34, the output unit 63 converts the time series data from the time series generation unit 62 into time series data corresponding to the processing target data and outputs it.

  According to the generation process using the time-series pattern storage network as described above, time-series data (of the time-series pattern) corresponding to the number of nodes in the time-series pattern storage network can be generated.

  As described above, the time series pattern storage network composed of a plurality of nodes having the time series pattern model 21 is updated in a self-organized manner based on the time series data. In addition to series data, unsupervised learning of time-series data whose length is not constant, that is, autonomous learning of time-series data can be easily (practically) performed.

  That is, each node of the time-series pattern storage network has the time-series pattern model 21. Therefore, in learning of the time-series pattern storage network, node update, that is, update of the time-series pattern model 21 possessed by the node is It does not affect the update of other nodes. Therefore, even if the number of nodes constituting the time series pattern storage network increases by one, the calculation amount required for learning of the time series pattern storage network simply increases by the calculation amount required for updating one node. Even if the scale of the time-series pattern storage network, that is, the number of nodes constituting the time-series pattern storage network is increased, the amount of calculation required for learning of the time-series pattern storage network does not increase dramatically. Therefore, even a large-scale time-series pattern storage network can be easily learned in a self-organizing manner.

  Furthermore, according to the learning of the time-series pattern storage network, the time-series pattern representing the statistical characteristics of the time-series data is stored in each node, so that the time-series data can be recognized using the time-series pattern. Generation can be done easily.

  Note that the learning process, the recognition process, and the generation process can be performed on, for example, audio data, image data, a signal for driving a motor (motor data), and other arbitrary time series data. Specifically, for example, the data processing apparatus of FIG. 10 is applied to an autonomous system such as an autonomous robot, and signals output from sensors corresponding to the vision, hearing, and touch of the robot, and the robot's hands and feet A signal that controls the motor that drives the part corresponding to the signal, a device that generates the synthesized sound, and a signal that is given to the LED that corresponds to the eye are used as time-series data for the learning, recognition, and generation processes. can do.

  In the present embodiment, an HMM, which is one of the state transition probability models, is adopted as the time series pattern model 21 (FIG. 7) possessed by the node. It is possible to employ a state transition probability model.

  As another state transition probability model that can be adopted as the time series pattern model 21, for example, there is a Bayesian network.

  In a Bayesian network, modeling is performed by expressing the dependency relationships between variables in a graph structure and assigning conditional probabilities to each node. In particular, by constructing a state transition model along the time axis, time series Data can be modeled.

  Note that the determination of the graph structure of the Bayesian network is performed, for example, by selecting a model that considers the likelihood of the learning data and the complexity of the graph structure. For example, the maximum likelihood estimation method is used to estimate the conditional probability. And EM (Expectation Maximaization) algorithm are used. Here, details of the Bayesian network are described in, for example, Yoichi Motomura, “Information Representation for Uncertainty Modeling: Bayesian Network”, 2001, Bayesian Network Tutorial.

  Furthermore, as the time series pattern model 21 (FIG. 7), as described above, a state transition probability model such as an HMM or a Bayesian network can be adopted, and a model approximating a function (hereinafter referred to as a function appropriately). It is also possible to adopt an approximate model).

  The function approximation model uses a time series pattern as a function f (). For example, a differential equation {x (t)} '= f (x (t)) or a difference equation x (t + 1) = f (x (t)) etc., and the function f () characterizes the time series pattern. Note that t represents time (time) (sample point), and x (t) represents a sample value of time-series data at time t or time-series data observed up to time (from 0) t. Further, {x (t)} ′ represents a first derivative with respect to time t of time series data x (t).

  Finding a function f () representing (corresponding to) a certain time series pattern from learning data (time series data) is called function approximation. For example, a function approximation method is a function using a polynomial or the like. There are a method of expressing f () and determining coefficients of the polynomial from learning data, a method of expressing a function f () by a neural network, and determining parameters of the neural network from learning data, and the like.

  In the function approximation of the function f () expressed by a polynomial, the coefficient of the polynomial can be determined (estimated) by, for example, the steepest descent method. Further, in the function approximation of the function f () expressed by the neural network, the parameters of the neural network can be determined by, for example, the back propagation method. Here, in the back-propagation method, input and output data are given to the neural network, and learning of the neural network parameters is performed so as to satisfy the relationship between the input and output data.

  For example, when a function approximation model that expresses a time series pattern with a differential equation x (t + 1) = f (x (t)) using a function f () is adopted as the time series pattern model 21, the input layer x A weight (intermediate layer) connecting (t) and the output layer x (t + 1) is a parameter of the neural network, and this parameter is learned using learning data (time-series data). The learning of the parameters of the neural network is performed by the back propagation method with appropriate initial values. As the neural network, for example, a recurrent neural network can be adopted.

  In learning of a time series pattern storage network composed of nodes having a time series pattern model 21 in which the function f () is expressed by a neural network, as in the case where the above-described HMM is adopted as the time series pattern model 21, It is necessary to determine a winner node, and in order to determine a winner node, it is necessary to calculate a score for new time-series data of each node of the time-series pattern storage network.

  As this score, for example, the observed value (actual value) of the new time series data and the theoretical value of the new time series data obtained from the time series pattern model 21 in which the function f () is expressed by a neural network. (For example, the sum of squared differences) can be employed. In this case, the node having the smallest score value is determined as the winner node that is the node most suitable for the new time-series data.

  After the winner node is determined, the update weight of each node is determined, and then the parameters of each node (the neural network) are updated in the same procedure as when the above-described HMM is adopted as the time series pattern model 21. be able to.

  For example, when the score is likelihood, the higher the score, the better the score. When the score is, for example, an error or distance, the smaller the score, the better the score.

  Next, a forward model and an inverse model to which the above time series pattern storage network is applied will be described.

  FIG. 20 schematically illustrates a configuration example of an input / output relationship model as a forward model or an inverse model to which the time-series pattern storage network is applied.

In FIG. 20, the input / output relationship model has two time-series pattern storage networks net _in and net _out . Furthermore, input-output relationship model when each node of the sequence pattern storage network net Non _in N _i and (i = 1, 2, · · ·, the total number of nodes), when each node N _'j of sequence pattern storage network net _out (J = 1, 2,..., The total number of nodes).

Here, coupling of the FIG. 20, the node N _i of the time series pattern storage network net Non _in, when a node N of the sequence pattern storage network net Non _out 'arrow between the _j is the node N _i and N' and _j Represents.

The time-series pattern storage networks net _in and net _out may have the same number of nodes and the same links (including the case where there is no link), or have different numbers of nodes or different links. It may be a thing. Also, when the time series pattern model 21 of the node N _i of sequence pattern storage network net Non _in (Fig. 7), the time series pattern model 21 which node has the time series pattern storage network net _out also, the same time series pattern model It may be a different time series pattern model.

  Next, FIG. 21 illustrates a configuration example of a data processing apparatus that performs various processes using an input / output relation model.

  In the data processing apparatus shown in FIG. 21, the control target is modeled as a forward model or an inverse model using the input / output relation model shown in FIG. Output data to be output and control data (input data) to be given to the controlled object are estimated.

  In other words, in FIG. 21, the data processing apparatus includes a storage unit 211, a learning unit 212, and a recognition generation unit 213.

  The storage unit 211 stores, for example, the input / output relationship model illustrated in FIG.

  The learning unit 212 includes teaching data that is a set of input data (observed values) given to the control target to be modeled and output data (observed values) obtained from the control target for the input data. Is to be supplied. Here, the teaching data is input data that is time-series data of a certain section (time-series data over a certain period of time), and other time-series data of a section obtained from the control target with respect to the time-series data of that section. A set with some output data.

  The learning unit 212 uses the teaching data supplied thereto to learn the input / output relationship model as a forward model or an inverse model to be controlled.

  That is, the learning unit 212 includes a learning processing unit 221 and a connection weight update unit 222.

The learning processing unit 221 is configured in the same manner as the learning processing unit 32 illustrated in FIG. 12, and based on input data of teaching data supplied to the learning unit 212, an input / output relation model stored in the storage unit 211. The time-series pattern storage network net _in in (FIG. 20) is updated in a self-organizing manner as in the case of the data processing apparatus in FIG. The learning processing unit 221 also inputs data stored in the storage unit 211 based on the output data (the output data set with the input data in the teaching data) of the teaching data supplied to the learning unit 212. The time-series pattern storage network net _out in the output relation model (FIG. 20) is updated in a self-organizing manner as in the case of the data processing apparatus of FIG.

Further, the learning processing unit 221, when the time to update the sequence pattern storage network net Non _in was the winning node, the time series pattern node label of the node N _i of the storage network net Non _in (hereinafter referred to as the input label) and , the time series pattern was the winning node to update the storage network net _out, the time series pattern storage network net _out node N _'j of node labels (hereinafter referred to as the output of the label) label set obtained by the set and Is supplied to the connection weight update unit 222.

Based on the label set supplied from the learning processing unit 221, the connection weight updating unit 222 is connected to the node N _i of the time-series pattern storage network net _{out in} the input / output relation model (FIG. 20) stored in the storage unit 211 and the time. The connection relationship with the node N ′ _j of the sequence pattern storage network net _out is updated.

Here, the label set supplied from the learning processing unit 221 to the connection weight updating unit 222 is a set of an input label and an output label, and the input label is stored in time series pattern based on the input data of the teaching data. was the winning node to update the network net Non _in, since it is the node label of the node N _i of the time series pattern storage network net Non _in, in the time series pattern storage network net Non _in, most compatible node input data N _i Is the node label.

Similarly, the output label is the node label of the node N ′ _j of the time-series pattern storage network net _out that became the winner node when updating the time-series pattern storage network net _out based on the output data of the teaching data. Therefore, in the time-series pattern storage network net _out , the node label of the node N ′ _j that best matches the output data.

In the weight updating unit 222, of the nodes of the time series pattern storage network net Non _in, and winning node N _i which is the most compatible node to the input data in the teaching data, with each node of the time series pattern storage network net _out While the connection relationship is updated, among the nodes of the time series pattern storage network net _out , the winner node N ′ _j that is the most suitable node for the output data in the teaching data, and each node of the time series pattern storage network net _in The connection relationship with is updated.

Here, the connection weight between the node of the time series pattern storage network net _{in and} the node of the time series pattern storage network net _out in the input / output relationship model is such that the stronger the degree of the connection, the larger the connection weight. The update of the connection relationship between the nodes means the update of the connection weight. Details of the method of updating the connection relationship between nodes by the connection weight update unit 222 will be described later.

  The recognition generator 213 receives input data for estimating output data obtained from the control target, or input data (control data) to be given to the control target in order to set the output data obtained from the control target as a certain target value. Output data for estimation is provided.

Then, the recognition generating unit 213, when the input data is supplied, the time series pattern storage network net Non _in the stored input-output relationship model in the storage unit 211, that best fits winning node N _i to the input data determined, the strongest coupling weight between the winning node N _i, the node of the time series pattern storage network net Non _out, the generating node N _'j of generating time-series data as the estimated value of the output data for the input data decide. Further, the recognition generation unit 213 generates and outputs output data (estimated value thereof) based on the time series pattern model 21 (FIG. 7) included in the generation node N ′ _j .

In addition, when the output data is supplied, the recognition generation unit 213 receives the winner node N ′ _j that best matches the output data in the time series pattern storage network net _out in the input / output relation model stored in the storage unit 211. Estimate the control data (input data) given to the control object when the output data is obtained for the node of the time-series pattern storage network net _{in that} has the strongest connection weight with the winner node N ' _j determining the generating node N _i for generating the time-series data as the value. Furthermore, the recognition generating unit 213, based on the time series pattern model 21 (FIG. 7) having the generating node N _i, and generates and outputs control data (estimated value of).

  That is, the recognition generation unit 213 includes a score calculation unit 231, a winner node determination unit 232, a generation node determination unit 233, and a time series generation unit 234.

  Note that the score calculation unit 231, the winner node determination unit 232, the generation node determination unit 233, and the time series generation unit 234 determine whether the data supplied to the recognition generation unit 213 is input data or output data. Can be recognized. That is, for example, for the recognition generation unit 213, whether the data is input data or output data separately from or together with the data supplied thereto. As a result, the score calculation unit 231, the winner node determination unit 232, the generation node determination unit 233, and the time series generation unit 234 receive the data supplied to the recognition generation unit 213. It recognizes whether it is input data or output data.

For the data supplied to the recognition generation unit 213, the score calculation unit 231 uses each node N _{i of} the time series pattern storage network net _in that constitutes the input / output relation model stored in the storage unit 211, or the time series pattern. A score, which is the degree to which each node N ′ _j of the storage network net _out matches, is calculated in the same manner as in the score calculation unit 51 of the recognition unit 3 in FIG. 16 and supplied to the winner node determination unit 232.

That is, when the data supplied to the recognition generation unit 213 is input data, the score calculation unit 231 has a time-series pattern storage network that constitutes an input / output relation model stored in the storage unit 211 for the input data. It calculates a score for each node N _i of net Non _in, supplied to the winning node determining unit 232. In addition, when the data supplied to the recognition generation unit 213 is output data, the score calculation unit 231 includes a time-series pattern storage network that forms an input / output relation model stored in the storage unit 211 for the output data. The score of each node N ′ _{j in} net _out is calculated and supplied to the winner node determination unit 232.

  Similarly to the case of the winner node determination unit 52 of the recognition unit 3 in FIG. 16, the winner node determination unit 232 determines the node having the highest score supplied from the score calculation unit 231 as the winner node, and represents the winner node. The node label is supplied to the generation node determination unit 233.

Therefore, when the data supplied to the recognition generation unit 213 is input data, the winner node determination unit 232 includes a node _in the time-series pattern storage network net _in that constitutes the input / output relationship model stored in the storage unit 211. in, it is supplied from the score computing unit 231, the highest node score for the input data is determined to winning node N _i, input label representing the winning node N _i is supplied to the generating node determining unit 233. In addition, when the data supplied to the recognition generation unit 213 is output data, the winner node determination unit 232 includes a node in the time-series pattern storage network net _out that constitutes the input / output relation model stored in the storage unit 211. in, is supplied from the score computing unit 231, the highest node score for output data, 'is determined to _j, the winning node N' winning node N output label representing a _j is supplied to the generating node determining unit 233 The

  Here, it is assumed that the highest (larger) score is the best score.

When the input label is supplied from the winner node determination unit 232, that is, when the data supplied to the recognition generation unit 213 is input data, the generation node determination unit 233 stores the input / output relation model stored in the storage unit 211. among the nodes of the time series pattern storage network net _out constituting the decision strongest link weight between the nodes N _i represented by the input label from the winning node determining unit 232 (strongest) node N _'j as a generation node Then, the output label representing the generation node N ′ _j is supplied to the time series generation unit 234. Also, the generation node determination unit 233 inputs / outputs stored in the storage unit 211 when the output label is supplied from the winner node determination unit 232, that is, when the data supplied to the recognition generation unit 213 is output data. Among the nodes of the time-series pattern storage network net _in constituting the relation model, the node N _i that generates the strongest (strongest) node _Ni with the connection weight with the node N ′ _j represented by the output label from the winner node determination unit 232 determined as to supply the input label representing the generating node N _i, the time-sequence generating unit 234.

When the output label is supplied from the generation node determination unit 233, that is, when the data supplied to the recognition generation unit 213 is input data, the time series generation unit 234 stores the input / output relation model stored in the storage unit 211. Based on the time-series pattern model 21 (FIG. 7) possessed by the node N ′ _j represented by the output label from the generation node determination unit 233 among the nodes of the time-series pattern storage network net _out that constitutes the recognition generation unit 213 The time series data as the estimated value of the output data with respect to the input data supplied to is generated, for example, in the same manner as in the time series generation unit 62 of the generation unit 6 in FIG.

Further, when the input label is supplied from the generation node determination unit 233, that is, when the data supplied to the recognition generation unit 213 is output data, the time series generation unit 234 inputs / outputs stored in the storage unit 211. based on the time series pattern storage network of the nodes of the net Non _in, time series pattern model 21 of the node N _i is represented by the input label from the generation node determining unit 233 constituting a related model, is supplied to the recognition generating unit 213 For example, the time series data as the estimated value of the control data (input data) for the output data is generated in the same manner as in the time series generation unit 62 of the generation unit 6 of FIG.

  Next, with reference to the flowchart of FIG. 22, modeling of a control target by the input / output relationship model, that is, learning of the input / output relationship model, which is performed in the data processing device of FIG.

  In step S <b> 101, when teaching data that is a set of input data and output data is input to the data processing apparatus of FIG. 21, the teaching data is supplied to the learning processing unit 221 of the learning unit 212.

In step S102, the learning processing unit 221 self-organizes the time series pattern storage network net _{in in} the input / output relation model (FIG. 20) stored in the storage unit 211 based on the input data of the teaching data. Update and go to step S103. In step S103, the learning processing unit 221 updates the time series pattern storage network net _out in the input / output relation model stored in the storage unit 211 in a self-organizing manner based on the output data of the teaching data.

Then, the learning processing unit 221, when time update sequence pattern storage network net Non _in was the winning node to an input label of the node N _i of the time series pattern storage network net Non _in, the time series pattern storage network net _out A label set that is a set with the output label of the node N ′ _j of the time-series pattern storage network net _out that became the winner node when updating is supplied to the connection weight update unit 222.

In step S104, the connection weight updating unit 222, based on the label set supplied from the learning processing unit 221, each node of the time-series pattern storage network net _{in in} the input / output relationship model (FIG. 20) stored in the storage unit 211. Then, the connection weight with each node of the time-series pattern storage network net _out is updated, the process returns to step S101, waits for the next teaching data to be input, and thereafter the same processing is repeated.

  By inputting a large number of teaching data and repeating the processing of steps S101 to S104, the input / output relationship model stored in the storage unit 211 becomes a forward model corresponding to the control target, and also an inverse model. It will become.

  Note that the processing of steps S102 and S103 can be performed in parallel, or can be performed in the reverse order of the case of FIG.

  Next, with reference to FIG. 23, the update of the connection weight (connection relationship between nodes) performed by the connection weight updating unit 222 (FIG. 21) in step S104 of FIG. 22 will be described.

Note that, hereinafter, as appropriate, in the input / output relation model, a time-series pattern storage network to which input data is given, that is, a time-series pattern model 21 that performs learning using the input data and expresses the time-series pattern of the input data. A time-series pattern storage network (for example, a time-series pattern storage network net _in ) composed of a plurality of nodes is called an input network, and learning is performed using a time-series pattern storage network to which output data is given, that is, output data. A time series pattern storage network (for example, a time series pattern storage network net _out ) composed of a plurality of nodes having a time series pattern model 21 expressing a time series pattern of output data is referred to as an output network.

  Further, hereinafter, a node of the input network is referred to as an input node, and a node of the output network is referred to as an output node as appropriate.

  In the following description, the input / output relation model is used as a forward model for estimating output data given input data, and the input / output relation model is given input data (target value). Description of the case of using it as an inverse model for estimating (control data) will be omitted as appropriate.

  FIG. 23 shows an input / output relationship model.

On the left side of FIG. 23, the input / output relationship model has one input network net _in and one output network net _out, and each input node of the input network net _in and each output node of the output network net _out Are connected. In FIG. 23, both the input network net _in and the output network net _out are configured by eight nodes.

On the left side of FIG. 23, each input node of the input network net _in is coupled to all output nodes of the output network net _out (thus each output node of the output network net _out is also all inputs of the input network net _in ). There is a connection weight w for every combination of input nodes of the input network net _in and output nodes of the output network net _out .

  Attention is now focused on two time-series pattern storage networks in which nodes are connected, and one of the time-series pattern storage networks is associated with each row, and the other time-series pattern storage network is connected to each column. And a matrix in which the connection weight w of the i-th node of one time-series pattern storage network and the j-th node of the other time-series pattern storage network is arranged in the element of the i-th row and j-th column. Assuming that the connection weight matrix MTX is used, the connection weight updating unit 222 (FIG. 21) updates the connection weight w that is each element of the connection weight matrix MTX.

  The right side of FIG. 23 shows the connection weight matrix MTX for the input / output relationship model on the left side of FIG.

  In the connection weight matrix MTX on the right side of FIG. 23, an input node is associated with each row, an output node is associated with each column, an i-th input node, and a j-th output node as elements of the i-th row and j-th column. And the connection weight w is arranged. The connection weight updating unit 222 (FIG. 21) updates the connection weight w that is each element of the connection weight matrix MTX.

That is, for example, when the power is first turned on, the connection weight updating unit 222 initializes all the connection weights w in the connection weight matrix MTX to, for example, 0 as an initial value. Then, the connection weight updating unit 222 receives teaching data, that is, a set of input data and output data, from the data processing apparatus shown in FIG. 21, and thereby an input label representing a winner node _in the input network net _in . Each time a label set with an output label representing a winner node in the output network net _out is given from the learning processing unit 221, the connection weight is updated with the connection between the winner nodes as the center.

  Specifically, the connection weight updating unit 222 updates the connection weight w of the connection weight matrix MTX, for example, according to the equation (4).

  Here, β is a learning rate representing the degree of updating the connection weight w, and is given in the range of 0 to 1. As the learning rate β decreases, the connection weight w does not change much. When the learning rate β is 0, the connection weight w does not change. On the other hand, as the learning rate β is increased, the connection weight w is also greatly changed. When the learning rate β is 1, the connection weight w is updated to the update reference value Δw.

  Further, the reference value Δw is given by, for example, Expression (5).

  Here, d represents the inter-pattern distance with the winner node, similarly to the case of Expression (3), and d = 0 for the node that is the winner node. Therefore, the reference value Δw is 1 for the winner node (the node that is), and the reference value Δw approaches 0 as the inter-pattern distance d from the winner node increases.

Now, the input node represented by the input label supplied from the learning processing unit 221 (FIG. 21) to the connection weight updating unit 222, that is, the winner node _in the input network net _in is represented as the input node _Ni and the learning processing unit 221. When the output node represented by the output label supplied from (FIG. 21) to the connection weight update unit 222, that is, the winner node in the output network net _out is represented as the output node N ′ _j , the connection weight update unit 222 (FIG. 21). Updates the connection weight w of the connection weight matrix MTX as follows according to Expression (4) (and Expression (5)).

That is, the weight updating unit 222, for each output node of the output network net Non _out, using the inter-pattern distance d between the output node N _'j is a winning node in the output network net Non _out, according to equation (5), the reference It obtains the value △ w, further the use of a reference value △ w, according to equation (4), and updates the weight w of the i-th input node N _i is the winning node of the input network net Non _in.

Thus, corresponding to the input node N _i is the winning node of the input network net Non _in, the i-th row of bond weight matrix MTX (each column) weight w is updated.

Further, the weight updating unit 222, for each input node of the input network net Non _in, using the inter-pattern distance d between the input node N _i is a winning node of the input network net Non _in, according to equation (5), the reference value Δw is obtained, and further, using the reference value Δw, the connection weight w with the _jth output node N ′ _j which is the winner node of the output network net _out is updated according to the equation (4).

As a result, the connection weight w (for each row) in the j-th column of the connection weight matrix MTX corresponding to the output node N ′ _j that is the winner node of the output network net _out is updated.

Therefore, the winning node N _i of the input network net Non _in, link weight between the winning node N _'j of the output network net _out is updated to enhance the most the degree of binding.

Note that the input node N _i is the winning node of the input network net Non _in, output update the weight w of the output node N _'j is the winning node of the network net _out is for each output node of the output network net _out , When updating the connection weight w with the input node N _i that is the winner node, or when updating the connection weight w with the output node N ′ _j that is the winner node for each input node of the input network net _in This is done only in either one.

  The updating of the connection weight w (connection weight matrix MTX) as described above is performed each time a set of input data and output data as teaching data is input to the data processing apparatus of FIG.

Furthermore, _in the learning based on the input data of the input network net _{in that} the input / output relation model has and the learning based on the output data of the output network net _out , the set of input data and output data as teaching data is the data shown in FIG. This is done each time it is input to the processing device.

When learning of the input network net _in and the output network net _out proceeds, the time series pattern model 21 possessed by the input node represents a specific time series pattern, and the time series pattern model 21 possessed by the output node. Will also express other specific time series patterns.

As a result, when there is some relationship between the input data of a specific time series pattern and the output data of another specific time series pattern, such a set of input data and output data (teaching data) ) Is given, the input node having the time series pattern model 21 representing a specific time series pattern _in the input network net _in becomes the winner node, and other specific time series patterns are expressed in the output network net _out . An output node having the time-series pattern model 21 to be a winner node.

Further, as described above, with the winner node of the input network net _in as the center, the connection weight of each input node of the input network net _{in and} the winner node of the output network net _out is updated, and the output network net The connection weight between each output node of the output network net _out and the winner node of the input network net _in is updated with the _out winner node as the center.

That is, the connection weight between each input node of the input network net _in and the winner node of the output network net _out seems to be stronger (enhanced) as the input distance between the patterns d near the winner node of the input network net _in is closer. Updated to Also, the connection weight between each output node of the output network net _out and the winner node of the input network net _in is updated so that the output node having a shorter inter-pattern distance d with the winner node of the output network net _out is updated. .

Conversely, each of the input nodes of the input network net Non _in, link weight between the winning node of the output network net _out the inter-pattern distance d between the winning node of the input network net Non _in becomes weaker the farther input node (weakening Updated). In addition, the connection weight between each output node of the output network net _out and the winner node of the input network net _in is also updated so that the output node with a longer pattern distance d from the winner node of the output network net _out becomes weaker. .

Given a lot of teaching data, the learning of the input network net _in and the output network net _out proceeds, and further, when the connection weight is updated, the input data (time series pattern) and the output data ( The input / output relation model as a forward model or an inverse model can be obtained.

According to the input / output relation model as a forward model, when certain input data is given, it is possible to determine the winner node that best matches the input data _in the input network net _in . The output node of the output network net _out having the strongest connection weight can be determined. Furthermore, by generating time-series data based on the output node (the time-series pattern model 21 that the node has), output data for given input data can be estimated.

Also, according to the input / output relationship model as an inverse model, when certain output data (target value) is given, the winner network that best matches the output data can be determined in the output network net _out . The input node of the input network net _in having the strongest connection weight with the winner node can be determined. Furthermore, control data (input data) for given output data can be estimated by generating time-series data based on the input node (a time-series pattern model 21 included in the input node).

  The connection weight matrix MTX is stored in the storage unit 211 (FIG. 21) as part of the input / output relationship model.

  Next, a process for estimating output data for input data and control data (input data) for output data using an input / output relation model as a forward model or an inverse model will be described with reference to the flowchart of FIG. .

  First, in the case of estimating output data for input data in the data processing device of FIG. 21, the input data is input to the data processing device of FIG. 21 in step S111.

  The input data input to the data processing apparatus is supplied to the score calculation unit 231 of the recognition generation unit 213.

Score calculating unit 231, in step S112, the input data supplied thereto, the score of each node N _i of the time series pattern storage network net Non _in configuring the input-output relationship model stored in the storage unit 211 Calculate and supply to the winner node determination unit 232.

In step S113, the winner node determination unit 232 selects the node having the highest score from the score calculation unit 231 among the nodes of the time-series pattern storage network net _in that constitute the input / output relationship model stored in the storage unit 211. determines the winning node N _i, provides input label representing the winning node N _i, the generating node determining unit 233.

In step S114, the generation node determination unit 233 represents the node represented by the input label from the winner node determination unit 232 among the nodes of the time-series pattern storage network net _out that constitute the input / output relationship model stored in the storage unit 211. A node N ′ _j having the strongest connection weight with N _i is determined as a generation node, and an output label representing the generation node N ′ _j is supplied to the time series generation unit 234.

In step S115, the time series generation unit 234 is a node represented by the output label from the generation node determination unit 233 among the nodes of the time series pattern storage network net _out that constitute the input / output relationship model stored in the storage unit 211. Based on the time-series pattern model 21 (FIG. 7) possessed by N ′ _j , time-series data is generated as an estimated value of output data for the input data supplied to the recognition generation unit 213, and the process proceeds to step S116. Output series data.

  Next, when the control data (input data) for the output data is estimated in the data processing apparatus of FIG. 21, the output data is input to the data processing apparatus of FIG. 21 in step S111.

  The output data input to the data processing apparatus is supplied to the score calculation unit 231 of the recognition generation unit 213.

In step S112, the score calculation unit 231 calculates the score of each node N ′ _j of the time-series pattern storage network net _out that constitutes the input / output relation model stored in the storage unit 211 for the output data supplied thereto. Is supplied to the winner node determination unit 232.

In step S113, the winner node determination unit 232 _selects the node having the highest score from the score calculation unit 231 among the nodes of the time-series pattern storage network net _out that constitute the input / output relationship model stored in the storage unit 211. The winner node N ′ _j is determined, and an output label representing the winner node N ′ _j is supplied to the generation node determination unit 233.

In step S114, the generation node determination unit 233 displays the node represented by the output label from the winner node determination unit 232 among the nodes of the time-series pattern storage network net _in that constitute the input / output relationship model stored in the storage unit 211. The node N _i having the strongest connection weight with N ′ _j is determined as a generation node, and an input label representing the generation node N _i is supplied to the time series generation unit 234.

In step S115, the time series generation unit 234 is a node represented by the input label from the generation node determination unit 233 among the nodes of the time series pattern storage network net _in that constitute the input / output relationship model stored in the storage unit 211. based on the time series pattern model 21 of the N _i, and generates time-series data as the estimated value of the control data (input data) to the output data supplied to the recognition generating unit 213, the process proceeds to step S116, the time series Output data.

  As described above, the input / output relation model is used to learn the input / output relationship model, and the controlled object is modeled into the forward model and the inverse model, so that the controlled object is expressed with high accuracy and high accuracy. It is possible to estimate output data and control data (input data).

  Next, as described above, in the data processing device of FIG. 21, input data that is time-series data (time-series data over a certain period of time) given to a control target and time-series data of that section. Are supplied to the learning unit 212 as teaching data, and a set of output data, which is other time-series data observed from the control target, is used in the learning unit 212. Learning the input / output relation model as a model or inverse model.

Therefore, for example, a time series of sensor data (voice data) detected by a microphone as a sensor is used as input data to be given to a robot as a control target, and the voice called to the robot When the learning of the input / output relation model is performed by adopting the time series of the motor data given to the motor in order to take the action of moving the arm so that the robot touches the hand, the learning unit 212 (FIG. 21) Self-organization so that one or more nodes of the input network net _in of the input / output relation model (the time series pattern model 21 possessed) express the time series pattern of the input data as the sensor data corresponding to the voice called to the robot. Learning and at least one node of the output network net _out Self-organizing learning is performed so that the column pattern model 21) expresses a time-series pattern of output data as motor data corresponding to the action of shaking hands.

Further, in the learning unit 212, the connection weight between each node of the input network net _in of the input / output relationship model and the winner node of the output network net _out is a node whose pattern distance d between the winner node of the input network net _in is short while being updated to be strong enough, output and each output node of the network net _out, link weight between the winning node of the input network net Non _in the output network net _out inter-pattern distance d is near the output node of the winning node of Updated to be stronger.

As a result, among the nodes of the input network net _in , one or more nodes (having the time series pattern model 21) representing the time series pattern of the input data as the sensor data corresponding to the voice called to the robot, and the output network Of the net _out nodes, the connection weight with one or more nodes (having the time series pattern model 21) representing the time series pattern of the motor data corresponding to the action of shaking hands becomes strong.

Thereafter, when sensor data corresponding to the voice calling the robot is given as input data to the recognition generation unit 213 (FIG. 21), the input / output relation model indicates that the input data is a node of the input network net _in . One of the nodes representing a time-series pattern of sensor data corresponding to the voice calling to the robot is a winner node. In the recognition generation unit 213, the motor corresponding to the action of waving among the nodes of the output network net _{out that} has the strongest connection weight with the winner node, that is, the nodes of the output network net _out in this case. A node that represents a time-series pattern of data is determined as a generation node, and output data, that is, motor data corresponding to the action of waving in this case, is generated and output using the generation node. The

Here, the recognition in the case of the robot's cognitive behavior using the input / output relation model corresponds to determining the winner node from the nodes of the input network net _in of the input / output relation model for the input data. The action corresponds to determining a generation node from among the nodes of the output network net _out for the winner node and generating time-series data (output data).

  According to the data processing apparatus of FIG. 21, for example, when a call voice is given to a robot, and the movement of an arm that shakes the hand is given to the voice, the learning as described above is performed, and as a result, Then, the robot will perform a cognitive action with a hand in response to the calling voice.

  However, in the data processing device of FIG. 21, it may be difficult to cause the robot to perform a task that requires real-time performance, which is completed by repeating recognition and action in a relatively short time. .

That is, for example, when the robot performs a task of rolling the ball in front of the right and left, there is a recognition to recognize (recognize) the state of the ball and an action to move the arm appropriately according to the state of the ball It is necessary to repeat it in a short predetermined time unit. Now, if the predetermined time, which is a unit of repetition, is called a frame, the sensor data in which a ball state is detected by a sensor in a certain frame on the time axis and the movement of the arm that must be taken in that frame 21 is provided as teaching data to the data processing apparatus of FIG. 21, and an input node (node of the input network net _in ) expressing a time-series pattern of sensor data in the teaching data; Even if the connection weight with the output node (node of the output network net _out ) that expresses the time-series pattern of the motor data in the teaching data is increased, the robot does the task of rolling the ball in front of it to the left and right. Will not be able to do.

  This is because when the sensor data of a certain frame is given, the movement of the arm based on the motor data of that frame must already be performed.

  Therefore, in order to allow the robot to perform the task of rolling the ball in front of the right and left, when attention is paid to a certain frame as the attention frame, of the sensor data in which the state of the ball is detected by the sensor, Of the input data representing the time-series pattern of the sensor data of the frame of interest (hereinafter also referred to as the node corresponding to the sensor data as appropriate) and the motor data of the motor data corresponding to the movement of the arm that rolls the ball left and right, In the input / output relationship model, the connection weight of the output node (hereinafter also referred to as the node corresponding to the motor data) expressing the time series pattern of the motor data of the frame delayed by a certain time from the time is strengthened. Learning is necessary.

  Depending on the recognition of the sensor data of the frame, that is, the processing speed of the process of determining the winner node for the sensor data of the frame, the motor data for the sensor data is generated after the sensor data of the frame of interest is given. There may be a delay time that cannot be ignored before the action is taken. The output node that increases the coupling weight with the node corresponding to the sensor data of the frame of interest is the delay corresponding to the motor data of the frame delayed from the frame of interest, if necessary. It is necessary to decide in consideration of time.

  FIG. 25 shows a configuration example of a robot capable of performing a task that requires real-time characteristics.

  In FIG. 25, the robot includes a data processing device 301, a sensor 302, a motor driving unit 303, and the like.

  The data processing device 301 performs self-organized learning of the input / output relation model using the time-series sensor data supplied from the sensor 302 and the time-series motor data supplied from the motor driving unit 303. Further, the data processing device 301 generates time-series motor data for the time-series sensor data supplied from the sensor 302 and supplies the time-series motor data to the motor driving unit 303.

  The sensor 302 is a camera, a microphone, or the like, detects an external state, and supplies time-series sensor data representing the external state to the data processing device 301.

  The motor driving unit 303 drives, for example, a motor (not shown) for moving a robot arm or the like according to the motor data supplied from the data processing device 301. The motor drive unit 303 is the same as the motor data to be given to the motor drive unit 303 in order to move, for example, when a robot arm or the like is moved by applying force from the outside. Motor data is generated and supplied to the data processing unit 301. Further, as described above, the motor driving unit 303 drives the motor in accordance with the motor data supplied from the data processing device 301, and supplies the motor data to the data processing device 301 as necessary. (return).

  25, the data processing device 301 includes a storage unit 311, a data extraction unit 312, a learning processing unit 313, a mapping learning unit 314, a data extraction unit 315, a recognition unit 316, a mapping unit 317, and a generation unit 318. Yes.

  The storage unit 311 stores an input / output relationship model.

Here, FIG. 26 shows the input / output relationship model M ₁₁ stored in the storage unit 311.

The input / output relation model M ₁₁ is a plurality of nodes having a time series pattern model that represents a time series pattern that is a pattern of time series data as input data, similarly to the input / output relation models shown in FIG. 20 and FIG. Input network net ₁ (input time series pattern storage network) which is a time series pattern storage network composed of a plurality of time series pattern models expressing time series patterns which are patterns of time series data as output data And an output network net ₂ (output time-series pattern storage network) which is a time-series pattern storage network composed of nodes.

In the input / output relationship model M ₁₁ , each node of the input network net ₁ and each node of the output network net ₂ are coupled by a coupling weight.

Returning to FIG. 25, sensor data output by the sensor 302 at each time and motor data output by the motor drive unit 302 at the same time are supplied to the data extraction unit 312 as teaching data. Data extracting unit 312, the sensor data of the teaching data, as input data to be given to the input-output relation model M ₁₁ during cognitive, from the time series of the input data, and sequentially extracts the input data for each frame, the learning processing To the unit 313.

Furthermore, the data extraction unit 312, the motor data of the teaching data, as output data to be generated from the input-output relationship model M ₁₁ during cognitive, from the time series of the output data sequentially output data for each frame Extracted and supplied to the learning processing unit 313.

The learning processing unit 313 includes a recognition learning processing unit 321 and a generation learning processing unit 322, and is stored in the storage unit 311 using input data and output data in units of frames supplied from the data extraction unit 312. Self-organized learning is performed for each of the input network net ₁ and the output network net _{2 included in the} input / output relationship model M ₁₁ (FIG. 26).

That is, the recognition learning processing unit 321 is similar to the learning processing unit 221 in FIG. 21, based on the input data in units of frames supplied from the data extraction unit 312, the input / output relationship model M ₁₁ stored in the storage unit 311. The input network net ₁ in (FIG. 26) is updated in a self-organizing manner.

Specifically, the recognition learning processing unit 321 obtains a score for the input data of the frame from the data extraction unit 312 for each node of the input network net ₁ in the input / output relationship model M ₁₁ , and determines the node of the input network net ₁ . The node with the best score is determined as the winner node (input winner node). Further, the recognition learning processing unit 321 updates the input network net ₁ in a self-organizing manner based on the input winner node for the input data of the frame from the data extraction unit 312.

  In addition, the recognition learning processing unit 321 supplies an input label, which is a node label representing the input winner node determined in time series for the input data in frame units, to the mapping learning unit 314 in time series.

Similarly to the learning processing unit 221 of FIG. 21, the generation learning processing unit 322 also uses the input / output relationship model M ₁₁ (FIG. ₁₁ ) stored in the storage unit 311 based on the output data in units of frames supplied from the data extraction unit 312. The output network net ₂ in 26) is updated in a self-organizing manner.

Specifically, the generation learning processing unit 322 obtains a score for the output data of the frame from the data extraction unit 312 for each node of the output network net ₂ in the input / output relationship model M ₁₁ , and determines the node of the output network net ₂ The node having the best score is determined as the winner node (output winner node). Further, the generation learning processing unit 322 updates the output network net ₂ in a self-organized manner based on the output winner node for the output data of the frame from the data extraction unit 312.

  Further, the generation learning processing unit 322 supplies an output label, which is a node label representing the output winner node determined in time series for the output data in frame units, to the mapping learning unit 314 in time series.

The mapping learning unit 314 includes a data buffer 331, a reading unit 332, and a coupling weight update unit 333, and inputs input data for each frame of the input network net ₁ in the input / output relation model M ₁₁ stored in the storage unit 311. and winning node updates the connection weights between each node in the output network net Non _2, an output winning node to the output of the network net Non _2, the output data for each frame delayed by a certain time from the time of the input data for each frame The connection weight with each node of the input network net ₁ is updated.

  That is, the data buffer 331 temporarily stores the time series input label supplied from the recognition learning processing unit 321 and the time series output label supplied from the generation learning processing unit 322.

  The reading unit 332 stores, in the data buffer 331, an input label representing an input winner node for the input data of the target frame, out of the input labels stored in the data buffer 331, as frames of the input data, sequentially as the target frame. Out of the output labels, the output label representing the output winner node corresponding to the output data of the frame delayed by a certain time from the time of the frame of interest is read out and associated, and the set of the associated input label and output label Is supplied to the connection weight update unit 333.

Similar to the connection weight update unit 222 in FIG. 21, the connection weight update unit 333 has the input / output relationship model stored in the storage unit 311 based on the label set supplied from the read unit 332 as described in FIG. and node N _i of the input network net Non ₁ in M ₁₁ (FIG. 26), the binding relationship between the node N _'j of the output network net Non _2, updates the Hebb rule or the like.

That is, the connection weight update unit 333 inputs the connection weight between each input node of the input network net ₁ and the output winner node of the output network net ₂ represented by the output label of the label set by the input network net represented by the input label of the label set. It updates the _first input winning node as the center, and respective output nodes of the output network net Non _2, the connection weight between the winning node of the input network net Non _1, updates about the winning node of the output network net Non _2.

Here, as described above, the input winner node is the winner node for the input data of the frame of interest, and the output winner node is the winner node for the output data of the frame delayed by a certain time from the frame of interest. according to the update of the connection weights by the weight updating unit 333, the input-output relationship model M ₁₁ stored in the storage unit 311, an input winning node of the input network net Non ₁ to the input data of the frame of interest, the constant from the target frame time The connection weight with the output winner node of the output network net ₂ with respect to the output data of the frame delayed by only is updated so as to become stronger.

As a result, the input-output relationship model M _11, the input data of a certain frame F is given, the input-output relationship model M _11, of the nodes of the input network net Non _1, node corresponding to the input data of the frame F Becomes the input winner node. Furthermore, the input-output relationship model M _11, of the nodes of the output network net Non _2, the node corresponding to the output data of a frame delayed by a fixed time from the frame F is the strongest link weight between the input winning node node As a result, the time series data corresponding to the output data of the frame delayed by a certain time from the frame F is generated based on the generation node.

  The sensor data output from the sensor 302 at each time is supplied to the data extraction unit 315. The data extraction unit 315 sequentially extracts the input data in frame units from the time series of the input data using the sensor data from the sensor 302 as input data, and supplies the extracted data to the recognition unit 316.

  The recognition unit 316 includes a score calculation unit 341 and a winner node determination unit 342.

  In FIG. 25, the mapping unit 317 includes a generation node determination unit 351. Furthermore, the generation unit 318 includes a time series generation unit 361.

  The score calculation unit 341, the winner node determination unit 342, the generation node determination unit 351, and the time series generation unit 361 are the score calculation unit 231, the winner node determination unit 232, the generation node determination unit 233, and the time series generation unit of FIG. Processing similar to that of H.234 is performed.

  The output data generated by the time series generation unit 361 is supplied to the motor driving unit 303 as motor data.

  Next, processing of the data extraction unit 312 in FIG. 25 will be described with reference to FIG.

  As described with reference to FIG. 25, the data extraction unit 312 is supplied with sensor data output by the sensor 302 at each time and motor data output by the motor driving unit 303 at the same time.

  Assuming that the length (time) of a frame is represented as T, the data extraction unit 312 uses sensor data as input data, and divides the time series of the input data into units of time T as shown in FIG. In addition, input data in units of frames is extracted and supplied to the learning processing unit 313.

  Further, the data extraction unit 312 extracts the output data in units of frames in which the motor data is output data, and the time series of the output data is divided into units of time T as shown in FIG. 27, and the learning processing unit 313. To supply.

  Here, hereinafter, a frame at time T having time t as the start time and time t + T as the end time will be referred to as a frame at time t as appropriate. If the input data at the time t (sample value) to the input data at the time t + t '(immediately before) is expressed as I [t, t + t'], the input data of the frame at the time t is , I [t, t + T]. Similarly, if the output data from time t to the output data at time t + t ′ is expressed as O [t, t + t ′], the output data of the frame at time t is O [t, t + T].

  Similarly to the data extraction unit 312, the data extraction unit 315 of FIG. 25 extracts input data in units of frames of time T from the time series of the input data using sensor data supplied from the sensor 302 as input data. And supplied to the recognition unit 316.

  Next, the processing of the reading unit 332 in FIG. 25 will be described with reference to FIG.

  As described above, the reading unit 332 sequentially sets the frames of the input data as the frame of interest, and among the input labels stored in the data buffer 331, the input label representing the input winner node for the input data of the frame of interest; Of the output labels stored in the data buffer 331, the output label representing the output winner node is associated with the output data of the frame delayed by a certain time from the time of the frame of interest.

That is, for example, when the same time T as the frame is used as the output data of the frame delayed by a certain time from the time of the frame of interest, the reading unit 332 reads the input data (I [t , t + T]) and the input label representing the input winner node N _t and the output data of the frame at time t + T delayed by time T from the time t of the frame of interest (O [t + T, t + 2T]) Is associated with an output label representing the output winner node N ′ _{t + T} for.

Next, with reference to the flowchart of FIG. 29, the learning processing which the robot performs the FIG. 25, i.e., it is described learning process of the input-output relationship model M _11.

  For example, if the robot learns the task of rolling the ball in front of it to the left or right, the operator (the user who wants the robot to learn the task) places the ball in front of the robot and holds the robot arm. And move the arm to roll the ball left and right.

  In this case, the sensor 302 detects the state of the ball rolling left and right, and time-series sensor data representing the state is supplied to the data extraction unit 312 of the data processing device 301.

  The motor driving unit 303 generates motor data corresponding to the movement of the arm being moved by the operator, and supplies the motor data to the data extraction unit 312 of the data processing unit 301.

  In step S301, the data extraction unit 312 uses the sensor data from the sensor 302 as input data, extracts input data in frame units from the time series of the input data, and supplies the input data to the recognition learning processing unit 321 of the learning processing unit 313. At the same time, the motor data from the motor drive unit 303 is used as output data, and output data in units of frames is extracted from the time series of the output data, and is supplied to the generation learning processing unit 322 of the learning processing unit 313, step S302. Proceed to

In step S302, the learning processing unit 313 uses the input data and output data in units of frames from the data extraction unit 312, and the input network net included in the input / output relationship model M ₁₁ (FIG. 26) stored in the storage unit 311. ₁ and self-organized learning of output network net ₂ .

That is, in step S302, the processing of steps S302 ₁ and S302 ₂ is performed.

In step S302 _1, the recognition learning processing unit 321, for each node of the input network net Non ₁ in the input-output relation model M ₁₁ stored in the storage unit 311, obtains a score for the input data frame from the data extraction unit 312 The node having the best score is determined as the input winner node from the nodes of the input network net ₁ . Further, the recognition learning processing unit 321 updates the input network net ₁ in a self-organized manner based on the input winner node for the input data in frame units from the data extraction unit 312, and displays an input label representing the input winner node. This is supplied to the mapping learning unit 314.

In step S302 _2, generating learning processing unit 322, for each node in the output network net Non ₂ in the input-output relation model M ₁₁ stored in the storage unit 311, obtains a score for output data of a frame unit from the data extraction unit 312 The node having the best score is determined as the output winner node from the nodes of the output network net ₂ . Furthermore, the generation learning processing unit 322 updates the output network net ₂ in a self-organized manner based on the output winner node for the output data in frame units from the data extraction unit 312, and outputs an output label representing the output winner node, This is supplied to the mapping learning unit 314.

  After the processing in step S302, the process proceeds to step S303, in which the mapping learning unit 314 uses, as an attention frame, a frame that has not yet been set as an attention frame among input data frames, and receives input data of the attention frame from the learning processing unit 313. An input label that represents an input winner node for is associated with an output label that represents an output winner node for output data of a frame delayed by time T from the frame of interest from the learning processing unit 313.

  That is, in step S303, in the data buffer 331 of the mapping learning unit 314, an input label indicating an input winner node for input data in units of frames from the learning processing unit 313 and an output label indicating an output winner node for output data in units of frames. Is temporarily stored.

  Further, in step S303, the reading unit 332 of the mapping learning unit 314 stores the input label representing the input winner node for the input data of the frame of interest among the input labels stored in the data buffer 331, and the data buffer 331. Among the output labels, the output label representing the output winner node corresponding to the output data of the frame delayed by time T from the time of the target frame is read and associated, and the associated input label and output label are set. The label set is supplied to the connection weight update unit 333.

Then, the process proceeds from step S303 to step S304, where the connection weight update unit 333 of the mapping learning unit 314 is based on the label set supplied from the reading unit 332 and the input / output relationship model M ₁₁ stored in the storage unit 311 (FIG. 26). ), The connection relationship between the node of the input network net _{1 and} the node of the output network net ₂ is updated.

That is, the connection weight updating unit 333 calculates the relationship between each input node of the input network net ₁ in the input / output relationship model M ₁₁ stored in the storage unit 311 and the output winner node of the output network net ₂ represented by the output label of the label set. The connection weight is updated centering on the input winner node of the input network net ₁ represented by the input label of the label set, and the connection weight of each output node of the output network net ₂ and the winner node of the input network net ₁ is output. Update around the winner node of network net ₂ .

  In the learning process of FIG. 29, the process from step S302 to S304 may be repeated by using the first frame to the last frame of the input data as the target frame in sequence, or the first frame of the input data. Steps S302 and S304 are repeated with the frame from the first frame to the last frame as the target frame in sequence, and then the process from step S303 and S304 is performed again sequentially from the first frame to the end frame of the input data as the target frame. May be repeated.

Next, with reference to the flowchart of FIG. 30, the processing of the cognitive behavior performed by the robot 25, i.e., generation processing (recognition generating processing) in time series using the input-output relationship model M ₁₁ will be explained.

  For example, as described with reference to FIG. 29, when the robot learns the task of rolling the ball in front of the eyes to the left and right and then puts the ball in front of the robot (and, if necessary, the ball is rolled ), The state of the ball is detected by the sensor 302, and time-series sensor data representing the state is supplied to the data extraction unit 315 of the data processing device 301.

  In step S311, the data extraction unit 315 uses the sensor data from the sensor 302 as input data, extracts input data in frame units from the time series of the input data, supplies the input data to the recognition unit 316, and proceeds to step S312. .

In the recognition unit 316, in step S <b> 312, the score calculation unit 341 sequentially uses the input data in units of frames from the data extraction unit 315 as the input data of the frame of interest, and the storage unit 311 for the input data of the frame of interest. The score of each node of the input network net ₁ constituting the input / output relationship model M ₁₁ stored in is calculated and supplied to the winner node determination unit 342.

Further, in step S312, the winning node determining unit 342, among the nodes of the input network net Non ₁ constituting the input-output relationship model M ₁₁ stored in the storage unit 311, the best score from the score computing unit 341 nodes To the input winner node, the input label representing the input winner node is supplied to the mapping unit 317, and the process proceeds to step S313.

In step S313, the generation node determination unit 351 of the mapping unit 317, among the nodes of the output network net Non ₂ constituting the input-output relationship model M ₁₁ stored in the storage unit 311, input label from the winning node determining unit 342 The node having the strongest connection weight with the node represented by (input winner node) is determined as the generation node, the output label indicating the generation node is supplied to the generation unit 318, and the process proceeds to step S314.

At step S314, the time series generating unit 361 of the generator 318, the output of the nodes of the network net Non _2, output label from the generation node determining unit 351 that constitutes the input-output relationship model M ₁₁ stored in the storage unit 311 Based on the time-series pattern model 21 (FIG. 7) of the generation node represented by, for example, the time-series data of the frame length is generated and output as the output data (estimated value) for the input data of the frame of interest. To do.

  The time series data as the output data is supplied to the motor drive unit 303, and the motor drive unit 303 drives the motor using the output data from the time series generation unit 361 as the motor data, thereby moving the robot arm. It is.

  Hereinafter, the processes in steps S312 to S314 are performed on the input data in units of frames supplied from the data extraction unit 315 to the recognition unit 316.

As described above, the input winning node to the input data for each frame of the input network net Non _1, updates the connection weights between each node in the output network net Non _2, the output of the network net Non _2, the time of the input data for each frame Since the connection weight between the output winner node for each frame of output data delayed by a certain time from each node and each node of the input network net ₁ is updated, the ball in front of the robot is rolled to the left and right. A task such as a task that requires real-time characteristics can be learned, and such a task can be performed.

  In the above case, the output data of the frame delayed by the same time T as the frame from the time of the frame of interest is adopted as the output data of the frame delayed by a certain time from the time of the frame of interest. That is, the time T is adopted as the time difference between the target frame of the input data and the frame of the output data corresponding to the target frame (hereinafter referred to as the input / output time difference as appropriate). For example, time 2T, 3T, 0.5T, or the like can be employed.

  As the input / output time difference, a time taking into account the time Δ required for the process of determining the winner node for the input data in units of frames, that is, for example, time T + Δ can be used.

  FIG. 31 shows input data and output data when the time T + Δ taking into account the time Δ required for the process of determining the winner node for the input data in units of frames is adopted as the input / output time difference.

In this case, output data corresponding to input data of the frame of interest, since the time from the time of the attention frame T + △ delayed frame of output data, the weight updating unit 333 (FIG. 25), the input network net Non ₁ The unit of frame that is updated by the time T + △ from the time of the input data of the frame unit of the output network net _{2 while} the connection weight of the input winner node for the input data of the frame unit and each node of the output network net ₂ is updated The connection weight between the output winner node for each output data and each node of the input network net ₁ is updated.

  In this case, the output data of the frame at time t + T + Δ corresponding to the input data I [t, t + T] of the frame at time t is represented as O [t + T + Δ, t + 2T + Δ]. The

  In addition, the data extraction unit 312 (FIG. 25) outputs, for example, output data in units of frames having a length of time T, which is delayed by a time Δ from the input data in units of frames having a length of time T. Must be extracted.

Here, the weight updating unit 333 (FIG. 25), an input winning node to the input data for each frame of the input network net Non _1, updates the connection weights between each node in the output network net Non _2, output network net Non ₂ Updating the connection weight between the output winner node for the frame unit output data delayed by the input / output time difference from the time of the frame unit input data and each node of the input network net ₁ as it associates the output data in units of frames delayed by output time difference from the time of the input data of the frame units, thereby performing learning of the input-output relationship model M _11.

Then, such learning using input-output relationship model M ₁₁ made, in the generation of output data to the input data, the input data for each frame, delayed input time difference from the time of the input data of the frame Output data for each frame (estimated value) is generated, and this output data is generated from the input data for each frame (for example, sensor data) by the input / output time difference from the time of the input data for each frame. It can be said that the output data of the future frame unit is predicted.

As described above, in the robot of FIG. 25, the input / output relationship model M _{11 is set} so that the input data in units of frames and the output data in units of frames delayed by the input / output time difference from the time of the input data in units of frames. To learn.

Therefore, when the robot's cognitive behavior (output data generation) matches the time series pattern expressed by any node of the input network net ₁ of the input / output relationship model M ₁₁ as sensor data that is input data. As long as series pattern sensor data (hereinafter referred to as known sensor data as appropriate) is input, motor data, which is output data appropriate to the sensor data, is generated. As a result, the robot performs the learned task. Can be reproduced.

  Next, experimental results of the experiment performed on the robot of FIG. 25 will be described.

  In the experiment, we adopted the task of rolling the ball with the arms of the left and right hands in front of the robot as a task for the robot to learn.

  In learning the task, we removed the gain of the arms that are the left and right hands of the robot, and reproduced the situation where a person rolls the ball left and right while operating the robot arm. Under the circumstances, information on the joint angles of the left and right arms of the robot is collected as motor data that is output data, and information on the position of the ball is collected as sensor data that is input data. A set of data and sensor data was used as teaching data for task learning.

  As the information on the position of the ball, a three-dimensional vector representing the coordinates (x, y, z) of the ball in the three-dimensional coordinate system defined on the space is adopted. An 8-dimensional vector representing the joint angle was adopted. Sensor data and motor data were sampled approximately every 70 ms.

  As the frame, 2100ms was used, and as sensor data for each frame, sample values for 2100ms were extracted from the sensor data with a 350ms offset while overlapping. The same applies to motor data in frame units.

  Note that sensor data and motor data in units of frames can be obtained by a person extracting sample values of a certain meaningful section from sensor data and motor data. However, since it is difficult to cause the robot to perform such extraction, in the experiment, as described above, sample values for 2100 ms are extracted with a shift of 350 ms, and sensor data and motor data in units of frames are extracted. did.

As the time series pattern storage network (input network net ₁ and output network net ₂ ) constituting the input / output relation model M ₁₁ , a time series pattern storage network having 10 × 10 nodes in the horizontal × vertical direction is adopted. Is provided with a link having a two-dimensional arrangement structure as shown in FIG.

  Further, an HMM is adopted as the time series pattern model 21 possessed by the node, and a single Gaussian (one Gaussian distribution) is adopted as the output probability density function of the HMM.

  In addition, since recognition of 2100ms of sensor data (processing to determine the winner node) can be performed in about one sample time at most, the winner node is determined for the sensor data in frame units for the input / output time difference. The time Δ required for the processing to be performed was not taken into consideration.

  In the experiment of cognitive behavior after learning, the movement of the arm tends to become somewhat unstable due to noise mixed in the sensor data, but when the ball starts to roll in front of the robot, the robot moves the ball left and right with the arm. It was confirmed to perform a rolling action. Therefore, it was verified that the robot of FIG. 25 can learn a task that requires real-time performance and can perform the task.

  Next, FIG. 32 shows another configuration example of a robot capable of performing a task that requires real-time characteristics.

  In the figure, portions corresponding to those of the robot of FIG. 25 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  In FIG. 32, the robot includes a data processing device 401, a sensor 302, a motor driving unit 303, and the like.

  The data processing device 401 performs self-organized learning of the input / output relation model using the time-series sensor data supplied from the sensor 302 and the time-series motor data supplied from the motor driving unit 303. Furthermore, the data processing device 401 generates time-series motor data for the time-series sensor data supplied from the sensor 302 and supplies the time-series motor data to the motor driving unit 303.

  32, the data processing device 401 includes a storage unit 411, a data extraction unit 312, a learning processing unit 413, a mapping learning unit 414, a data extraction unit 315, a recognition unit 316, a mapping unit 417, and a generation unit 318. Yes.

  The storage unit 411 stores an input / output relationship model.

Here, FIG. 33 shows the input / output relationship model M ₁₂₃ stored in the storage unit 411.

The input / output relationship model _M123 has a plurality of time-series pattern storage networks, and a node of one time-series pattern storage network and a node of another time-series storage pattern network are coupled by a coupling weight. So common with input-output relationship model M ₁₁ in FIG. 26.

However, the input / output relationship model M ₁₁ in FIG. 26 has two input network net ₁ and output network net ₂ as time series pattern storage networks, and the input network net ₁ node and the output network net ₂ The input / output relationship model M ₁₂₃ of FIG. 33 has three time-series pattern storage networks, one of which is the time-series pattern storage network node and the other one of the time series. FIG. 26 shows that the node of the pattern storage network is coupled, and the other one of the time-series pattern storage network nodes and the other one of the time-series pattern storage network nodes are coupled. It is different from the output relation model M _11.

In other words, the input / output relationship model M ₁₂₃ includes two time series pattern storage networks net ₁ composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as input data. net ₃ (first and second input time series pattern storage network) and a time series composed of a plurality of nodes having a time series pattern model expressing a time series pattern which is a pattern of time series data as output data Pattern storage network net ₂ (output time-series pattern storage network).

Here, two time series pattern storage networks net ₁ and net ₃ composed of a plurality of nodes having a time series pattern model expressing a time series pattern of input data are respectively set to a first input network net ₁ and a second input network. Input network called net ₃ . Similar to the input-output relationship model M ₁₁ in FIG. 26, the time series pattern storage network net Non ₂ including a plurality of nodes having a time series pattern model representing a time series pattern of the output data, the output network net Non ₂ That's it.

In the input / output relationship model M ₁₂₃ , the node of the first input network net _{1 and} the node of the second input network net ₃ are coupled by the coupling weight (w ₁₃ ), and the second input network net ₃ and the node of net ₂ of the output network are coupled by the coupling weight (w ₃₂ ).

  Returning to FIG. 32, the learning processing unit 413 includes a recognition learning processing unit 321 and a generation learning processing unit 322, similarly to the learning processing unit 313 of FIG. 25. Further, the learning processing unit 413 includes a recognition learning unit 421 in addition to the recognition learning processing unit 321 and the generation learning processing unit 322.

  The recognition learning unit 421 performs the same processing as the recognition learning processing unit 321.

That is, input data in units of frames is supplied from the data extraction unit 312 to the recognition learning unit 421, similarly to the recognition learning unit 321. Similar to the recognition learning unit 321, the recognition learning unit 421, based on the input data in units of frames supplied from the data extraction unit 312, stores the input / output relationship model M ₁₂₃ stored in the storage unit 411 (FIG. 33). The second input network net ₃ at is updated in a self-organizing manner.

Specifically, the recognition learning processing unit 421 obtains a score for the input data in frame units from the data extraction unit 312 for each node of the second input network net ₃ in the input / output relationship model M ₁₂₃ , and the second The node having the best score among the nodes of the input network net ₃ is determined as the winner node (input winner node). Further, the recognition learning processing unit 421 updates the second input network net ₃ in a self-organized manner based on the input winner node for the input data in frame units from the data extraction unit 312.

  In addition, the recognition learning processing unit 421 supplies an input label, which is a node label representing the input winner node determined in time series with respect to the input data in frame units, to the mapping learning unit 414 in time series.

Here, as described with reference to FIG. 25, the recognition learning processing unit 321 scores for the input data in units of frames from the data extraction unit 312 for each node of the first input network net ₁ in the input / output relationship model M ₁₂₃ . And the node having the best score among the nodes of the first input network net ₁ is determined as the winner node (input winner node). In order to distinguish between the input winner node determined from the nodes of the first input network net ₁ and the input winner node determined from the nodes of the second input network net ₃ described above, The input winner node determined from the nodes of the first input network net ₁ is referred to as a first input winner node, and the input winner node determined from the nodes of the second input network net ₃ is This is called the second input winner node.

  Furthermore, hereinafter, a node label representing the first input winner node will be referred to as a first input label, and a node label representing the second input winner node will be referred to as a second input label as appropriate.

The mapping learning unit 414 includes a data buffer 331, a reading unit 432, and a coupling weight update unit 433, and a certain frame F of the first input network net ₁ in the input / output relationship model M ₁₂₃ stored in the storage unit 411. The connection weight (w ₁₃ ) between the first input winner node for each input data and each node of the second input network net ₃ is updated, and the input data of the frame F of the second input network net ₃ is updated. The connection weight between the second input winner node for the input data of the frame delayed by the input / output time difference from the time and each node of the _first input network net ₁ is updated.

Further, the mapping learning unit 414 includes a second input winner node for the input data of the second input network net ₃ with respect to the input data of the frame delayed by the input / output time difference from the time of the input data of the frame F, and each of the output networks net ₂ updates the connection weights between the nodes, the output of the network net Non _2, an output winning node to the output data of a frame delayed by output time difference from the time of the input data of the frame F, each node of the second input network net Non ₃ Update the connection weight with.

  That is, in the mapping learning unit 414, the data buffer 331 temporarily stores the first input label, the second input label, and the output label supplied from the learning processing unit 413 in time series.

  That is, in the embodiment of FIG. 32, in the learning processing unit 413, the recognition learning processing unit 321 supplies the first input label to the mapping learning unit 414, and the recognition learning processing unit 421 maps the second input label. Since the generation learning processing unit 322 supplies the output label to the mapping learning unit 414, the data buffer 331 of the mapping learning unit 414 supplies these first input label, second input label, And the output label is stored.

  The reading unit 432 sequentially sets the frames of the input data as the frames of interest, and among the first input labels stored in the data buffer 331, the first input label representing the input winner node for the input data of the frames of interest The second input label representing the second input winner node for the input data of the frame delayed by the input / output time difference from the time of the frame of interest among the second input labels stored in the data buffer 331 is read out. The association and the label set (hereinafter, appropriately referred to as the input label set) that is a set of the associated first input label and second input label are supplied to the connection weight update unit 433.

  Further, the reading unit 432 represents a second input winner node representing the second input winner node for the input data of the frame delayed by the input / output time difference from the time of the frame of interest among the second input labels stored in the data buffer 331. Read and associate the input label with the output label representing the output winner node for the output data of the frame delayed by the input / output time difference from the time of the frame of interest among the output labels stored in the data buffer 331, and the association In addition, a label set (hereinafter, referred to as an input / output label set as appropriate) that is a set of the second input label and the output label is supplied to the connection weight update unit 433.

Similar to the connection weight update unit 222 in FIG. 21, the connection weight update unit 433 performs input / output stored in the storage unit 411 based on the input label set supplied from the reading unit 432 as described in FIG. 23. The connection relationship between the node of the first input network net _{1 and} the node of the second input network net _{3 in} the relationship model M ₁₂₃ (FIG. 33) is updated according to the Hebb rule or the like. Further, the connection weight updating unit 433 is configured to output the second input network net ₃ node and the output network in the input / output relation model M ₁₂₃ stored in the storage unit 411 based on the input / output label set supplied from the reading unit 432. The connection relationship with the net ₂ node is updated according to the Hebb rule.

That is, the connection weight updating unit 433 combines each input node of the first input network net ₁ and the second input winner node of the second input network net ₃ represented by the second input label of the input label set. The weight (w ₁₃ ) is updated around the first input winner node of the first input network net ₁ represented by the first input label of the input label set, and each input node of the second input network net ₃ And the first connection weight for updating the connection weight with the first input winner node of the first input network net _{1 around} the second input winner node of the second input network net ₃ is updated. .

Further, the connection weight updating unit 433 inputs the connection weight (w ₃₂ ) between each input node of the second input network net ₃ and the output winner node of the output network net ₂ represented by the output label of the input / output label set. Update the second input winner net node of the second input network net ₃ represented by the second input label of the output label set, and update each output node of the output network net ₂ and the second input network net ₃ The second connection weight is updated to update the connection weight with the second input winner node around the output winner node of the output network net ₂ .

  The mapping unit 417 includes a generation node determination unit 451, determines a generation node that generates time-series data based on the node label representing the winner node supplied from the recognition unit 316, and represents the generation node The node label is supplied to the generation unit 318.

That is, the generation node determination unit 451 receives from the winner node determination unit 342 of the recognition unit 316 a node label (first label node representing the first winner node for input data in frame units among the nodes of the _first input network net _1. Input label) is supplied.

The generation node determination unit 451 selects the node having the strongest combination with the first input winner node represented by the node label from the winner node determination unit 342 among the nodes of the second input network net ₃ (hereinafter, the strongest as appropriate). In addition, the strongest coupling node and the node having the strongest coupling among the nodes of the output network net ₂ are determined as generation nodes. Then, the generation node determination unit 451 supplies a node label (output label) representing the generation node to the generation unit 318.

  Next, processing of the reading unit 432 in FIG. 32 will be described with reference to FIG.

  As described above, the reading unit 432 uses the first input winner node for the input data of the frame of interest among the first input labels stored in the data buffer 331 as the frame of the input data sequentially as the frame of interest. Of the first input label representing the second input label stored in the data buffer 331, the second input winner node representing the second input winner node for the input data of the frame delayed by the input / output time difference from the time of the frame of interest 2 input labels are associated with each other.

  Further, the reading unit 432 represents a second input winner node representing the second input winner node for the input data of the frame delayed by the input / output time difference from the time of the frame of interest among the second input labels stored in the data buffer 331. The input label is associated with the output label representing the output winner node for the output data of the frame delayed by the input / output time difference from the time of the frame of interest among the output labels stored in the data buffer 331.

That is, for example, when the same time T as the frame is adopted as the input / output time difference, the reading unit 432 reads the first input winner for the input data (I [t, t + T]) of the frame at time t. A second input for the first input label representing the node N _t and the input data (I [t + T, t + 2T]) of the frame at time t + T delayed by the input / output time difference T from the time t of the frame of interest A second input label representing the input winner node N _{t + T} is associated.

Further, the reading unit 432 outputs the second input winner node N _t for the input data (I [t + T, t + 2T]) of the frame at time t + T delayed by the input / output time difference T from the time t of the frame of interest. Output winner node N for the second input label representing _{+ T} and the output data (O [t + T, t + 2T]) of the frame at time t + T delayed by the input / output time difference T from time t of the frame of interest 'Associate with output label representing _{t + T.}

Next, the learning process performed by the robot of FIG. 32, that is, the learning process of the input / output relationship model _M123 will be described with reference to the flowchart of FIG.

  For example, if the robot learns the task of rolling the ball in front of it to the left or right, the operator moves the arm to place the ball in front of the robot, hold the robot arm, and roll the ball to the left or right. .

  In this case, the sensor 302 detects the state of the ball rolling left and right, and time-series sensor data representing the state is supplied to the data extraction unit 312 of the data processing device 401.

  Further, the motor drive unit 303 generates motor data corresponding to the movement of the arm being moved by the operator, and supplies the motor data to the data extraction unit 312 of the data processing device 401.

  In step S331, the data extraction unit 312 uses the sensor data from the sensor 302 as input data, extracts input data in units of frames from the time series of the input data, and recognizes learning processing units 321 and 421 of the learning processing unit 413. , The motor data from the motor drive unit 303 is used as output data, output data in frame units is extracted from the time series of the output data, and is supplied to the generation learning processing unit 322 of the learning processing unit 413. Proceed to step S332.

In step S332, the learning processing unit 413 uses the input data and output data in units of frames from the data extraction unit 312, and the first input / output relationship model M ₁₂₃ (FIG. 33) stored in the storage unit 411 has. Self-organized learning of the input network net ₁ , the second input network net ₃ , and the output network net ₂ is performed.

That is, in step S332, the processes of steps S332 ₁ , S332 ₂ , and S332 ₃ are performed.

In step S332 ₁ , the recognition learning processing unit 321 applies the frame unit input data from the data extraction unit 312 for each node of the first input network net ₁ in the input / output relationship model M ₁₂₃ stored in the storage unit 411. The score is obtained, and the node having the best score among the nodes of the first input network net ₁ is determined as the first input winner node. Further, the recognition learning processing unit 321 updates the first input network net ₁ in a self-organized manner based on the first input winner node for the input data in units of frames from the data extraction unit 312, The first input label representing the input winner node is supplied to the mapping learning unit 414.

In step S < _b > 332 ₂ , the recognition learning processing unit 421 applies the frame unit input data from the data extraction unit 312 for each node of the second input network net ₃ in the input / output relationship model M ₁₂₃ stored in the storage unit 411. The score is obtained, and the node having the best score is determined as the second input winner node from the nodes of the second input network net ₃ . Furthermore, the recognition learning processing unit 421 updates the second input network net ₃ in a self-organized manner based on the second input winner node for the input data in frame units from the data extraction unit 312, A second input label representing the input winner node is supplied to the mapping learning unit 414.

In step S332 ₃ , the generation learning processing unit 322 obtains a score for the output data in units of frames from the data extraction unit 312 for each node of the output network net ₂ in the input / output relationship model M ₁₂₃ stored in the storage unit 411. The node having the best score is determined as the output winner node from the nodes of the output network net ₂ . Furthermore, the generation learning processing unit 322 updates the output network net ₂ in a self-organized manner based on the output winner node for the output data in frame units from the data extraction unit 312, and outputs an output label representing the output winner node, This is supplied to the mapping learning unit 414.

  After the processing of step S332, the process proceeds to step S333, and the mapping learning unit 414 uses the input data frame of the input data from the learning processing unit 413 as a target frame, which has not been set as the target frame. The first input label representing the first input winner node for the second input and the second input representing the second input winner node for the input data of the frame delayed by the input / output time difference T from the frame of interest from the learning processing unit 413 A second input label representing a second input winner node for the input data of the frame delayed from the target frame by the input / output time difference T from the learning processing unit 413, and the learning processing unit 413, Shows the output winner node for the output data of the frame delayed by I / O time difference T from the frame of interest Correspondence between the output label.

  That is, in step S333, in the data buffer 331 of the mapping learning unit 414, the first and second input labels respectively representing the first and second input winner nodes for the input data in frame units from the learning processing unit 413, and An output label representing an output winner node for output data in frame units is temporarily stored.

In step S333, the read section 432 of the picture learning unit 414 performs the process of step S333 ₁ and S333 _2.

That is, in step S333 _1, the read unit 432, a first input label representing one of the first input label stored in the data buffer 331, a first input winning node to the input data of the frame of interest, the data Of the second input labels stored in the buffer 331, the second input label representing the second input winner node corresponding to the input data of the frame delayed by the input / output time difference T from the time of the target frame is read and corresponded The input label set, which is a set of the first and second input labels associated with each other, is supplied to the connection weight updating unit 433.

In step S333 _2, reading unit 432 of the second input label stored in the data buffer 331, a second input winning node to the input data of a frame delayed by output time difference T from the time frame of interest The second input label to be represented and the output label representing the output winner node for the output data of the frame delayed by the input / output time difference T from the time of the frame of interest among the output labels stored in the data buffer 331 are read and corresponded The input / output label set, which is a set of the associated second input label and output label, is supplied to the connection weight updating unit 433.

Then, the process proceeds from step S333 to step S334, and the connection weight update unit 433 of the mapping learning unit 414 stores the input / output relationship stored in the storage unit 411 based on the input label set and the input / output label set supplied from the reading unit 432. Update of the first connection weight for updating the connection relationship between the node of the first input network net _{1 and} the node of the second input network net _{3 in} the model M ₁₂₃ (FIG. 33), and the second input network net The second connection weight is updated to update the connection relationship between the node _{3 and} the node of the output network net ₂ .

That is, the connection weight update unit 433 updates each of the input nodes of the first input network net ₁ in the input / output relationship model M ₁₂₃ stored in the storage unit 411 and the input label set as the update of the first connection weight. The first input winner of the first input network net ₁ represented by the first input label of the input label set represents the connection weight with the second input winner node of the second input network net ₃ represented by the two input labels. And updating the connection weights of the respective input nodes of the second input network net ₃ and the first input winner node of the first input network net ₁ in the second input network net ₃ Update with 2 input winner nodes as the center.

Further, the connection weight updating unit 433 updates each input node of the second input network net ₃ in the input / output relationship model M ₁₂₃ stored in the storage unit 411 and the input / output label set as the second connection weight update. The connection weight with the output winner node of the output network net ₂ represented by the output label is updated around the second input winner node of the second input network net ₃ represented by the second input label of the input / output label set. The connection weight between each output node of the output network net ₂ and the second input winner node of the second input network net ₃ is updated with the output winner node of the output network net ₂ as the center.

According to the learning process of the input / output relationship model M ₁₂₃ as described above, the connection relationship between the first input network net ₁ and the second input network net _{3 included in} the input / output relationship model M ₁₂₃ (FIG. 33), and The connection relationship between the second input network net ₃ and the output network net ₂ is updated as follows.

That is, the first input winner node is a winner node for the input data of the frame of interest, and the second input winner node is a winner node for the input data of the frame delayed by the input / output time difference T from the frame of interest. according by the weight updating unit 433 to the first connection weight updating, the input-output relationship model M ₁₂₃ stored in the storage unit 411, first the first input of the input network net Non ₁ to the input data of the frame of interest The connection weight (w ₁₃ ) between the winner node and the second input winner node of the second input network net ₃ for the input data of the frame delayed by the input / output time difference T from the frame of interest is updated to be stronger. The

Furthermore, since the output winner node is a winner node for the output data of the frame delayed by the input / output time difference T from the frame of interest, when the second connection weight is updated by the connection weight update unit 433, the output winner node is stored in the storage unit 411. in the input-output relationship model M ₁₂₃ which is a second input winning node of the second input network net Non ₃ to the input data of a frame delayed by output time difference T from the frame of interest, only input and output time difference T from the target frame delay The connection weight (w ₃₂ ) with the output winner node of the output network net ₂ for the output data of the frame is updated to be stronger.

  In the learning process of FIG. 35, the process from step S332 to S334 may be repeated with the frame from the first frame to the last frame of the input data as the target frame, or the first frame of the input data. Steps S332 and S334 are repeatedly performed with the frame from the first frame to the last frame as the target frame, and then the process from step S333 and S334 is performed again with the first frame to the last frame of the input data as the target frame again. May be repeated.

Next, a process of cognitive behavior performed by the robot of FIG. 32, that is, a process of generating time-series data (recognition generation process) using the input / output relationship model _M123 will be described with reference to the flowchart of FIG.

  For example, as described with reference to FIG. 35, when the robot learns the task of rolling the ball in front of the eyes to the left and right, and then puts the ball in front of the robot (and if necessary, the ball is rolled. ), The state of the ball is detected by the sensor 302, and time-series sensor data representing the state is supplied to the data extraction unit 315 of the data processing device 401.

  In step S341, the data extraction unit 315 uses the sensor data from the sensor 302 as input data, extracts input data in frame units from the time series of the input data, supplies the input data to the recognition unit 316, and proceeds to step S342. .

In the recognition unit 316, in step S342, the score calculation unit 341 sequentially uses the input data in units of frames from the data extraction unit 315 as the input data of the frame of interest, and the storage unit 411 for the input data of the frame of interest. The score of each node of the first input network net ₁ constituting the input / output relation model M ₁₂₃ (FIG. 33) stored in is calculated and supplied to the winner node determination unit 342.

Further, in step S342, the winner node determination unit 342 has the highest score from the score calculation unit 341 among the nodes of the input network net ₁ constituting the input / output relationship model M ₁₂₃ stored in the storage unit 411. Is determined as the first input winner node, and the first input label representing the first input winner node is supplied to the mapping unit 417, and the process proceeds to step S343.

In step S343, the generation node determination unit 451 of the mapping unit 417 selects from the winner node determination unit 342 among the nodes of the second input network net ₃ constituting the input / output relationship model M ₁₂₃ stored in the storage unit 411. The node having the strongest connection weight with the node represented by the first input label (first input winner node) is obtained as the strongest connection node. Further, in step S343, the generation node determination unit 451 determines a node having the strongest connection weight with the strongest connection node of the second input network net ₃ among the nodes of the output network net ₂ as the generation node. The output label representing the generation node is supplied to the generation unit 318, and the process proceeds to step S344.

In step S344, the time series generating unit 361 of the generator 318, the output of the nodes of the network net Non _2, output label from the generation node determining unit 451 included in a storage unit 411 input-output relationship model M ₁₂₃ stored in Based on the time-series pattern model 21 (FIG. 7) of the generation node represented by, for example, the time-series data of the frame length is generated and output as the output data (estimated value) for the input data of the frame of interest. To do.

  The time series data as the output data is supplied to the motor drive unit 303, and the motor drive unit 303 drives the motor using the output data from the time series generation unit 361 as the motor data, thereby moving the robot arm. It is.

  Thereafter, the processes in steps S342 to S344 are performed on the input data in units of frames supplied from the data extraction unit 315 to the recognition unit 316.

As described above, in the robot shown in FIG. 32, the first input winner node for the input data in the frame unit of the first input network net ₁ and the time of the input data in the frame unit of the second input network net ₃ are used. The input data of the frame unit of the second input winner node and the output network net ₂ are updated so as to increase the connection weight with the second input winner node for the input data of the frame unit delayed by the input / output time difference. 25, the output weight of the frame unit delayed by the input / output time difference is updated so as to increase the connection weight with the output winner node. Therefore, as in the case of the robot of FIG. It is possible to learn a task that requires real-time characteristics, such as a task that rolls to a point, and perform such a task.

That is, according to the learning process of FIG. 35 as described above, in the input / output relationship model M ₁₂₃ stored in the storage unit 411, the first input winner node of the first input network net ₁ for the input data of the frame of interest. And the connection weight (w ₁₃ ) with the second input winner node of the second input network net ₃ for the input data of the frame delayed by the input / output time difference from the frame of interest is updated to be stronger, Output of the second input network net ₃ for the input data of the frame delayed by the input / output time difference from the target frame and the output of the output network net ₂ for the output data of the frame delayed by the input / output time difference from the target frame The connection weight (w ₃₂ ) with the winner node is updated to be stronger.

Therefore, in recognition generating process of FIG. 36, the input-output relationship model M _123, the input data of a certain frame F is given, the input-output relationship model M _123, one of the first input network net Non ₁ node, The node corresponding to the input data of frame F becomes the first input winner node. Further, in the input / output relationship model _M123 , the node corresponding to the input data of the frame delayed by the input / output time difference from the frame F among the nodes of the _second input network net ₂ is the first input winner node. The node having the strongest connection weight is determined as the strongest connection node.

36, the node corresponding to the output data of the frame delayed by the input / output time difference from the frame F among the nodes of the output network net ₂ is the node having the strongest connection weight with the strongest connection node. Thus, it is determined as a generation node, whereby time-series data corresponding to the output data of the frame delayed by a certain time from the frame F is generated based on the generation node.

  Therefore, in the robot of FIG. 32, sensor data as input data in units of frames, that is, motor data as output data of frames delayed by an input / output time difference from input data in units of frames with respect to the external state, that is, Since learning is performed so that motor data corresponding to the action to be taken can be obtained, learning a task that requires real-time characteristics such as a task of rolling the ball in front of the eyes to the left and right, You will be able to perform such tasks.

Here, the update of the first connection weight by the connection weight update unit 433 (FIG. 32), that is, the first input winner node for the sensor data as the input data in units of frames of the first input network net ₁ , In the second input network net ₃ , updating the frame unit sensor data delayed by the input / output time difference from the frame unit sensor data time so as to increase the connection weight with the second input winner node sensor data, to associate the sensor data for each frame delayed by output time difference from the time of the sensor data in the frame unit, thereby performing learning of the input-output relationship model M _123.

Also, the second connection weight is updated by the connection weight updating unit 433, that is, the second sensor data in the second input network net ₃ is delayed by the input / output time difference from the time of the sensor data in the frame unit. Updating the input winner node and the output network net ₂ to increase the connection weight between the output winner node for the motor data as output data in frame units delayed by the input / output time difference from the time of sensor data in frame units. The input / output relationship model _M123 is learned so that the sensor data in units of frames and the sensor data in units of frames and motor data in units of frames at the same time are associated with each other.

When sensor data in units of frames is given to the input / output relationship model _M123 in which such learning has been performed at the time of cognitive behavior, in the input / output relationship model _M123 , for the sensor data in units of frames, the frame The node corresponding to the sensor data (estimated value) of the frame unit delayed by the input / output time difference from the time of the sensor data of the unit is obtained as the strongest coupling node. It can be said that sensor data corresponding to the strongest coupling node, that is, sensor data of a future frame is predicted from the time of frame F by an input / output time difference.

Furthermore, at the time of cognitive behavior, the node with the strongest connection weight with the strongest connection node among the nodes of the output network net ₂ , that is, the sensor data of the frame unit delayed by the input / output time difference from the time of the sensor data of the frame unit. A node corresponding to the motor data (estimated value) of the frame unit delayed by the input / output time difference is obtained as a generation node via the prediction, but this is obtained from the sensor data of a certain frame F, It can be said that the motor data corresponding to the generation node, that is, the motor data of the future frame is predicted by the input / output time difference from the time of frame F.

As described above, in the robot of FIG. 32, the input / output relationship is set so that the input data I in units of frames and the input data I ′ in units of frames delayed by an input / output time difference from the time of the input data in units of frames. In addition to learning the model _M123 , input data I ′ in frame units that are delayed by an input / output time difference from the time of input data I in frame units, and frame units that are delayed by an input / output time difference from the time of input data in frames The input / output relationship model _M123 is learned so as to be associated with the output data O ′.

For this reason, at the time of the robot's cognitive behavior (output data generation), as the sensor data that is the input data, the time-series pattern expressed by any node of the first input network net ₁ of the input / output relationship model M ₁₂₃ is used. As long as the matching time-series pattern sensor data (known sensor data) is input, motor data that is appropriate output data is generated for the sensor data, and as a result, the robot reproduces the learned task. can do.

  Next, the experimental results of the experiment performed on the robot of FIG. 32 will be described.

  For the robot of FIG. 32, an experiment similar to that performed for the robot of FIG. 25 described above is performed. As a result, the movement of the arm tends to become somewhat unstable due to noise mixed in the sensor data. When the ball started to roll, it was confirmed that the robot performed the action of rolling the ball left and right with the arm. Therefore, it was verified that the robot shown in FIG. 32 can learn a task that requires real-time performance and can perform the task.

From the above, according to the input / output relationship model _M123, when sensor data of a frame at time t is given, sensor data of a future frame by the input / output time difference T from time t, that is, at time t + T The sensor data of the frame can be predicted. The sensor data of a future frame can be predicted by repeatedly giving the predicted sensor data to the input / output relationship model _M123 . Prediction value of the sensor data, depending on whether the sensor 302 when the task is learning is performed in the input-output relationship model M ₁₂₃ corresponds to the external state detecting, the storage structure of the input-output relationship model M ₁₂₃ It is possible to check whether or not the learned task (a time series pattern of sensor data representing an external state observed when the task is performed) is appropriately stored.

32, the first input network net ₁ and the second input network net ₂ included in the input / output relationship model M ₁₂₃ stored in the storage unit 411 have the input / output relationship model in FIG. Similar to the time-series pattern storage networks net _in and net _out , the number of nodes, the links, and the time-series pattern models 21 included in the nodes may be the same or different.

On the other hand, in the robot of FIG. 32, both the first input network net ₁ and the second input network net _{2 included in} the input / output relationship model M ₁₂₃ stored in the storage unit 411 are input data in units of frames. It is updated in a self-organized manner using the sensor data.

Accordingly, the first input network net Non ₁ and as a second input network net Non _2, in the case of adopting the same time series pattern storage network, input network net Non ₂ of the first input network net Non ₁ and the second is One time series pattern storage network can be substituted.

That is, FIG. 37 shows an input / output relationship model M ′ in which the first input network net ₁ and the second input network net _{2 included} in the input / output relationship model M ₁₂₃ of FIG. ₁₂₃ is shown.

The input / output relationship model M ′ _{123 in} FIG. 37 has an input network net ₁ and an output network net ₂ as two time-series pattern storage networks, and the first input / output relationship model M _{123 in} FIG. The input network net ₁ and the second input network net ₂ are substituted by the input network net ₁ .

In the input / output relationship model M ′ _{123 of} FIG. 37, the input network net ₁ is substituted for the first input network net ₁ and the second input network net ₂ that the input / output relationship model M ₁₂₃ of FIG. Therefore, the nodes of the input network net ₁ are coupled by the coupling weight (w ₁₃ ), and the node of the input network net _{1 and} the node of the output network net ₂ are coupled by the coupling weight (w ₃₂ ). Are bound by.

FIG. 38 shows a configuration example of a robot that performs task learning and cognitive behavior using the input / output relationship model M ′ ₁₂₃ of FIG.

  In the figure, parts corresponding to those of the robot of FIG. 25 or FIG. 32 are given the same reference numerals, and description thereof will be omitted below as appropriate.

  That is, the robot shown in FIG. 38 replaces the storage unit 411, the reading unit 432, the combination weight update unit 433, and the generation node determination unit 451 shown in FIG. 32, respectively, with a storage unit 1411, a read unit 1432, a combination weight update unit 1433, A generation node determination unit 1451 is provided, and the configuration is the same as that of the robot of FIG. 32 except that the recognition learning processing unit 421 of FIG. 32 is not provided.

The storage unit 1411 stores the input / output relationship model M ′ ₁₂₃ shown in FIG.

The reading unit 1432 sequentially stores the frames of the input data as the frames of interest, the input labels stored in the data buffer 331, that is, the storage unit 1411 performed by the recognition learning processing unit 321 using the input data in units of frames. Of the input labels representing the input winner node obtained in the self-organizing update of the input network net ₁ of the stored input / output relationship model M ′ ₁₂₃ (FIG. 37), this represents the input winner node for the input data of the frame of interest. An input label (this input label corresponds to the above-mentioned “first input label”, and is hereinafter referred to as a first input label as appropriate) and input data of a frame delayed by an input / output time difference from the time of the frame of interest An input label representing an input winner node for (this input label corresponds to the above-mentioned “second input label”. (Referred to as “second input label” as appropriate), and a label set (input label set) that is a set of the first input label and the second input label that are associated with each other is set as a connection weight update unit. 1433.

  Further, the reading unit 1432 includes a second input label representing an input winner node for input data of a frame that is delayed by an input / output time difference from the time of the frame of interest among the input labels stored in the data buffer 331, and a data buffer Among the output labels stored in 331, the output label representing the output winner node corresponding to the output data of the frame delayed by the input / output time difference from the time of the target frame is read and associated, and the associated second input label A label set (input / output label set) that is a set of the output label and the output label is supplied to the connection weight update unit 1433.

Here, an input winner node of the input network net ₁ in the input / output relationship model M ′ ₁₂₃ (input winner node represented by the first input label) obtained for the input data of the target frame with a certain frame as the target frame. Corresponds to the above-mentioned “first input winner node”, and is hereinafter referred to as a first input winner node as appropriate. The input winner node of the input network net ₁ in the input / output relation model M ′ ₁₂₃ (input winner node represented by the second input label) obtained for the input data of the frame delayed by the input / output time difference from the time of the frame of interest ) Corresponds to the above-mentioned “second input winner node”, and is hereinafter referred to as a second input winner node as appropriate.

Similar to the connection weight update unit 222 in FIG. 21, the connection weight update unit 1433 is the input / output stored in the storage unit 1411 based on the input label set supplied from the read unit 1432 as described in FIG. 23. The connection relationship between the nodes of the input network net _{1 in} the relationship model M ′ ₁₂₃ (FIG. 37) is updated by the Hebb rule or the like. Further, the connection weight updating unit 1433 is based on the input / output label set supplied from the reading unit 1432, the node of the input network net ₁ in the input / output relationship model M ′ ₁₂₃ stored in the storage unit 1411, and the output network net _2. The connection relationship with the node is updated according to the Hebb rule or the like.

That is, the weight updating unit 1433, and the input node of the input network net Non _1, the coupling weight between the second input winning node of the input network net Non ₁ second input label of the input label set represents (w ₁₃₎ , and updates about the first input winning node of the input network net Non ₁ representing the first input label of the input label set, and each input node of the input network net Non _1, a first input winner of the input network net Non ₁ The connection weight with the node is updated around the second input winner node of the input network net ₁ . Note that this update of the connection weight corresponds to the above-described “update of the first connection weight”, and hence is hereinafter referred to as update of the first connection weight as appropriate.

Further, the connection weight update unit 1433 sets the connection weight (w ₃₂ ) between each input node of the input network net ₁ and the output winner node of the output network net ₂ represented by the output label of the input / output label set. The second input winner node of the input network net ₁ represented by the second input label of the input network is updated, and each output node of the output network net ₂ is combined with the second input winner node of the input network net ₁ The weight is updated around the output winner node of the output network net ₂ . Note that this update of the connection weight corresponds to the above-described “update of the second connection weight”, and hence is hereinafter referred to as update of the second connection weight as appropriate.

  The generation node determination unit 1451 determines a generation node that generates time-series data based on the node label that represents the winner node supplied from the recognition unit 316, and generates a node label that represents the generation node. To supply.

That is, the generation node determination unit 1451 receives from the winner node determination unit 342 of the recognition unit 316 a node label (input label) representing the first winner node for input data in frame units among the nodes of the input network net _1. Is to be supplied.

The generation node determination unit 1451 is the node having the strongest combination with the first input winner node represented by the node label from the winner node determination unit 342 among the nodes of the input network net ₁ (this node is also referred to as appropriate below). Further, the node having the strongest connection node and the strongest connection among the nodes of the output network net ₂ is determined as the generation node. Then, the generation node determination unit 1451 supplies the generation unit 318 with a node label (output label) representing the generation node.

Next, the learning process performed by the robot of FIG. 38, that is, the learning process of the input / output relationship model M ′ ₁₂₃ will be described with reference to the flowchart of FIG.

  For example, if the robot learns the task of rolling the ball in front of it to the left or right, the operator moves the arm to place the ball in front of the robot, hold the robot arm, and roll the ball to the left or right. .

  In this case, the sensor 302 detects the state of the ball rolling left and right, and time-series sensor data representing the state is supplied to the data extraction unit 312 of the data processing device 401.

  Further, the motor drive unit 303 generates motor data corresponding to the movement of the arm being moved by the operator, and supplies the motor data to the data extraction unit 312 of the data processing device 401.

  In step S1331, the data extraction unit 312 extracts sensor data from the sensor 302 as input data, extracts input data in frame units from the time series of the input data, and supplies the input data to the recognition learning processing unit 321 of the learning processing unit 413. At the same time, the motor data from the motor drive unit 303 is used as output data, and output data in units of frames is extracted from the time series of the output data, and is supplied to the generation learning processing unit 322 of the learning processing unit 413, step S1332 Proceed to

In step S1332, the learning processing unit 413 uses the input data and output data in units of frames from the data extraction unit 312, and the input network included in the input / output relationship model M ′ ₁₂₃ (FIG. 37) stored in the storage unit 1411. Self-organized learning of net ₁ and output network net ₂ .

That is, in step S1332, the processing in steps S1332 ₁ and S1332 ₂ is performed.

In step S1332 _1, the recognition learning processing unit 321, for each node of the input network net Non ₁ in the input-output relation model M _'123 stored in the storage unit 1411, a score for the input data frame from the data extraction unit 312 The node having the highest score is determined as the input winner node from the nodes of the input network net ₁ . Further, the recognition learning processing unit 321 updates the input network net ₁ in a self-organized manner based on the input winner node for the input data in frame units from the data extraction unit 312, and displays an input label representing the input winner node. This is supplied to the mapping learning unit 414.

In step S1332 _2, generating learning processing unit 322, for each node in the output network net Non ₂ in the input-output relation model M _'123 stored in the storage unit 1411, a score for output data of a frame unit from the data extraction unit 312 The node having the highest score is determined as the output winner node from the nodes of the output network net ₂ . Furthermore, the generation learning processing unit 322 updates the output network net ₂ in a self-organized manner based on the output winner node for the output data in frame units from the data extraction unit 312, and outputs an output label representing the output winner node, This is supplied to the mapping learning unit 414.

  After the processing of step S1332, the process proceeds to step S1333, and the mapping learning unit 414 uses the frames of the input data that have not yet been set as the target frame as the target frame, and receives the input data of the target frame from the learning processing unit 413. The winner node for the input data of the first input label representing the first input winner node and the input data of the frame delayed from the target frame by the input / output time difference T from the learning frame 413, that is, the second And a second input label representing the second input winner node, and a frame delayed from the target frame by the input / output time difference T from the learning processing unit 413. The output label representing the output winner node for the output data is associated.

  That is, in step S 1333, in the data buffer 331 of the mapping learning unit 414, an input label representing an input winner node for input data in units of frames from the learning processing unit 413 and an output label representing an output winner node for output data in units of frames. Is temporarily stored.

In step S 1333, the reading unit 1432 of the mapping learning unit 414 performs the processes in steps S 1333 ₁ and S 1333 ₂ .

That is, in step S1333 _1, reading unit 1432 of the input label stored in the data buffer 331, a first input label representing the first input winning node to the input data of the frame of interest, the time frame of interest The second input label representing the second input winner node is read and associated with the input data of the frame delayed by the input / output time difference T from the input, and the input is a set of the associated first and second input labels The label set is supplied to the connection weight update unit 1433.

In step S13332, the reading unit 1432 displays the _second input winner node representing the second input winner node for the input data of the frame delayed by the input / output time difference T from the time of the frame of interest among the input labels stored in the data buffer 331. And the output label representing the output winner node for the output data of the frame delayed by the input / output time difference T from the time of the frame of interest among the output labels stored in the data buffer 331, An input / output label set that is a set of the associated second input label and output label is supplied to the connection weight update unit 1433.

Then, the process proceeds from step S 1333 to step S 1334, and the connection weight update unit 1433 of the mapping learning unit 414 stores the input / output relationship stored in the storage unit 1411 based on the input label set and the input / output label set supplied from the reading unit 1432. The update of the first connection weight for updating the connection relationship between the nodes of the input network net ₁ in the model M ′ ₁₂₃ (FIG. 37), and the connection relationship between the node of the input network net _{1 and} the node of the output network net ₂ The second connection weight to be updated is updated.

That is, the connection weight update unit 1433 updates each input node of the input network net ₁ in the input / output relationship model M ′ ₁₂₃ stored in the storage unit 1411 and the second input label set as the first connection weight update. the coupling weight between the second input winning node of the input network net Non ₁ the input label representing, with updates about the first input winning node of the input network net Non ₁ representing the first input label of the input label set, each input node of the input network net Non _1, and updates the connection weights between the first input winning node of the input network net Non _1, about a second input winning node of the input network net Non _1.

Further, the connection weight update unit 1433 updates each input node of the input network net ₁ in the input / output relationship model M ′ ₁₂₃ stored in the storage unit 1411 and the output label of the input / output label set as the second connection weight update. connection weights between the output winning node of the output network net Non ₂ represented, along with updates about a second input winning node of the input network net Non ₁ second input label of the input and output label set represents, output network net Non ₂ And the connection weight of the second input winner node of the input network net ₁ is updated with the output winner node of the output network net ₂ as the center.

According to the learning process of the input / output relationship model M ′ ₁₂₃ as described above, the connection relationship between the nodes of the input network net _{1 included in} the input / output relationship model M ′ ₁₂₃ (FIG. 37), and the second input network net ₃ And the connection relationship between the nodes of the output network net ₂ is updated as follows.

That is, the first input winner node is a winner node for the input data of the frame of interest, and the second input winner node is a winner node for the input data of the frame delayed by the input / output time difference T from the frame of interest. according by the weight updating unit 1433 in the first connection weight updating, the input-output relationship model M _'123 stored in the storage unit 1411, a first input winning node of the input network net Non ₁ to the input data of the frame of interest Then, the connection weight (w ₁₃ ) with the second input winner node of the input network net ₁ for the input data of the frame delayed by the input / output time difference T from the frame of interest is updated so as to be stronger.

Further, since the output winner node is a winner node for the output data of the frame delayed by the input / output time difference T from the target frame, the second connection weight is updated by the connection weight update unit 1433 and stored in the storage unit 1411. In the input / output relationship model M ′ ₁₂₃ , the second input winner node of the input network net ₁ for the input data of the frame delayed by the input / output time difference T from the target frame, and the frame delayed by the input / output time difference T from the target frame The connection weight (w ₃₂ ) with the output winner node of the output network net ₂ for the output data is updated so as to be stronger.

  Note that in the learning process of FIG. 39, the processes from step S1332 to S1334 may be repeated with the first frame to the last frame of the input data as the target frame, or the first frame of the input data. The process from step S1332 is repeated with the frame from the first frame to the end frame as the target frame in sequence, and then the processes from step S1333 and S1334 are performed again sequentially from the first frame to the end frame of the input data as the target frame. May be repeated.

Next, a process of cognitive behavior performed by the robot of FIG. 38, that is, a process of generating time-series data (recognition generating process) using the input / output relation model M ′ ₁₂₃ will be described with reference to the flowchart of FIG.

  For example, as described with reference to FIG. 39, when the robot learns the task of rolling the ball in front of the eyes to the left and right and then puts the ball in front of the robot (and, if necessary, the ball is rolled ), The state of the ball is detected by the sensor 302, and time-series sensor data representing the state is supplied to the data extraction unit 315 of the data processing device 401.

  In step S1341, the data extraction unit 315 uses the sensor data from the sensor 302 as input data, extracts input data in units of frames from the time series of the input data, supplies the input data to the recognition unit 316, and proceeds to step S1342. .

In the recognition unit 316, in step S1342, the score calculation unit 341 sequentially uses the input data in units of frames from the data extraction unit 315 as the input data of the frame of interest, and the storage unit 1411 for the input data of the frame of interest. The score of each node of the input network net ₁ constituting the input / output relation model M ′ ₁₂₃ (FIG. 37) stored in is calculated and supplied to the winner node determination unit 342.

Further, in step S1342, the winner node determination unit 342 has the highest score from the score calculation unit 341 among the nodes of the input network net ₁ that constitute the input / output relationship model M ′ ₁₂₃ stored in the storage unit 1411. The node is determined to be an input winner node, and an input label representing the input winner node is supplied to the mapping unit 417, and the process proceeds to step S1343.

In step S 1343, the generation node determination unit 1451 of the mapping unit 417 receives the input from the winner node determination unit 342 among the nodes of the input network net ₁ constituting the input / output relationship model M ′ ₁₂₃ stored in the storage unit 1411. The node having the strongest connection weight with the node represented by the label (input winner node) is obtained as the strongest connection node. Further, in step S1343, the generation node determination unit 1451 determines a node having the strongest connection weight with the strongest connection node of the input network net ₁ among the nodes of the output network net ₂ , as the generation node. The output label to represent is supplied to the production | generation part 318, and it progresses to step S1344.

In step S 1344, the time series generation unit 361 of the generation unit 318 outputs the output from the generation node determination unit 1451 among the nodes of the output network net ₂ that constitute the input / output relationship model M ′ ₁₂₃ stored in the storage unit 1411. Based on the time series pattern model 21 (FIG. 7) of the generation node represented by the label, for example, time series data of the frame length is generated as output data (estimated value) for the input data of the frame of interest. Output.

  The time series data as the output data is supplied to the motor drive unit 303, and the motor drive unit 303 drives the motor using the output data from the time series generation unit 361 as the motor data, thereby moving the robot arm. It is.

  Thereafter, the processes in steps S1342 to S1344 are performed on the input data in units of frames supplied from the data extraction unit 315 to the recognition unit 316.

As described above, in the robot shown in FIG. 38, the first input winner node for the input data in frame units of the input network net ₁ and the input data in frame units delayed by the input / output time difference from the time of the input data in frame units. Frame unit delayed by input / output time difference from the time of the input data of the frame unit between the second input winner node and the output network net ₂ As shown in FIG. 32, for example, a real-time property such as a task of rolling the ball in front of the right and left is required. Will be able to learn and perform such tasks.

Further, in the robot shown in FIG. 38, the first input network net ₁ and the second input network net _{2 included} in the input / output relationship model M ₁₂₃ shown in FIG. 33 used by the robot shown in FIG. Since the input / output relationship model M ′ ₁₂₃ (FIG. 37) substituted for the network is used, the storage capacity required to store one time-series pattern storage network (second input network net ₂ ) Can be saved.

  Next, FIG. 41 shows still another configuration example of a robot capable of performing a task that requires real-time characteristics.

  In the figure, portions corresponding to those of the robot of FIG. 25 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  In FIG. 41, the robot includes a data processing device 501, a sensor 302, a motor drive unit 303, and the like.

  The data processing device 501 performs self-organized learning of the input / output relationship model using the time-series sensor data supplied from the sensor 302 and the time-series motor data supplied from the motor driving unit 303. Further, the data processing device 501 generates time-series sensor data and motor data for the time-series sensor data supplied from the sensor 302 and the time-series motor data supplied from the motor drive unit 303. The motor data is supplied to the motor drive unit 303.

  That is, in FIG. 41, the data processing device 501 is replaced with the storage unit 311 and the data extraction unit 312 of FIG. 25, respectively, except that a storage unit 511 and a data extraction unit 512 are provided. The configuration is the same as 301.

  The storage unit 511 stores an input / output relationship model.

Here, FIG. 42 shows the input / output relationship model M ₁₁₁₂ stored in the storage unit 511.

The input / output relation model M ₁₁₁₂ is composed of a plurality of nodes having a time series pattern model that represents a time series pattern that is a pattern of time series data, similarly to the input / output relation model shown in FIGS. An input network net ₁₁ (input time series pattern storage network) which is a time series pattern storage network and an output network net ₁₂ (output time series pattern storage network) are provided.

In the input / output relationship model M ₁₁₁₂ , each node of the input network net ₁₁ and each node of the output network net ₁₂ are coupled by a coupling weight.

  Returning to FIG. 41, sensor data output by the sensor 302 at each time and motor data output by the motor driving unit 302 at the same time are supplied to the data extraction unit 512 as teaching data. The data extraction unit 512 sequentially extracts the input data in units of frames from the time series of the input data as the input data and the output data as a time series of vectors having the sensor data and the motor data as the teaching data as components, Output data in units of frames are sequentially extracted from the time series of output data, and the input data and output data in units of frames are supplied to the learning processing unit 313.

  That is, FIG. 43 shows input data and output data handled by the data extraction unit 512 of FIG.

  As described above, the input data and the output data handled by the data extraction unit 512 are both time series of vectors having sensor data and motor data as components, and are the same time series data.

  As described in FIG. 41, the data extraction unit 512 is supplied with sensor data output by the sensor 302 at each time and motor data output by the motor driving unit 303 at the same time.

  For example, as described above, if the frame length (time) is represented by T, the data extraction unit 512 uses the sensor data and motor data as input data, and the time series of the input data is shown in FIG. As described above, input data in units of frames divided in units of time T is extracted and supplied to the learning processing unit 313.

  Further, the data extraction unit 512 extracts sensor data and motor data as output data, extracts output data in units of frames by dividing the time series of the output data into units of time T as shown in FIG. The data is supplied to the processing unit 313.

  Here, IO [t, t + t from the time t vector (sample value) to the vector at time t + t '(immediately before) in the time series of vectors having sensor data and motor data as components. If expressed as'], both the input data and output data of the frame at time t can be expressed as IO [t, t + T].

  In FIG. 43, the input data and the output data are completely the same time series data. However, the input data and the output data may be partially the same time series data. That is, for example, the input data is a vector time series including sensor data and motor data as components, and the output data is a vector time series including motor data (and time series data other than sensor data) as components. It may be.

  Next, the processing of the reading unit 332 in FIG. 41 will be described with reference to FIG.

  As described with reference to FIGS. 25 and 28, the reading unit 332 sequentially sets the input data frames as the target frames, and among the input labels stored in the data buffer 331, the input winner node for the input data of the target frames. Is associated with an output label representing an output winner node for output data of a frame delayed by an input / output time difference from the time of the frame of interest among the output labels stored in the data buffer 331.

That is, for example, when the same time T as the frame is used as the input / output time difference, the reading unit 332 reads the input winner node N _t for the input data (IO [t, t + T]) of the frame at time _t. And an output representing the output winner node N ′ _{t + T} for the output data (IO [t + T, t + 2T]) of the frame at time t + T delayed by time T from the time t of the frame of interest Associate a label.

  However, here, the input data and the output data are the same time-series data (a vector time-series having sensor data and motor data as components).

Next, the learning process performed by the robot of FIG. 41, that is, the learning process of the input / output relationship model _M1112 will be described with reference to the flowchart of FIG.

  For example, if the robot learns the task of rolling the ball in front of it to the left or right, the operator moves the arm to place the ball in front of the robot, hold the robot arm, and roll the ball to the left or right. .

  In this case, the sensor 302 detects the state of the ball that rolls left and right, and time-series sensor data representing the state is supplied to the data extraction unit 512 of the data processing device 501.

  In the motor drive unit 303, motor data corresponding to the movement of the arm being moved by the operator is generated and supplied to the data extraction unit 512 of the data processing device 501.

  In step S361, the data extraction unit 512 uses, as input data, a time series of vectors (hereinafter, referred to as sensor motor data as appropriate) having the sensor data from the sensor 302 and the motor data from the motor driving unit 303 as components. The input data in frame units is extracted from the time series of the input data, supplied to the recognition learning processing unit 321 of the learning processing unit 313, and the sensor motor data as output data, from the time series of the output data in frame units. Is output to the generation learning processing unit 322 of the learning processing unit 313, and the process proceeds to step S362.

In step S362, the learning processing unit 313 uses the input data and output data in units of frames from the data extraction unit 512, and the input network net included in the input / output relationship model M ₁₁₁₂ (FIG. 42) stored in the storage unit 511. ₁₁ and output network net ₁₂ self-organized learning.

That is, in step S362, the processes of steps S362 ₁ and S362 ₂ are performed.

In step S362 _1, the recognition learning processing unit 321, for each node of the input network net Non ₁₁ in the input-output relation model M ₁₁₁₂ stored in the storage unit 511, obtains a score for the input data frame from the data extraction unit 512 The node having the best score among the nodes of the input network net ₁₁ is determined as the input winner node. Further, the recognition learning processing unit 321 updates the input network net ₁₁ in a self-organized manner based on the input winner node for the input data in frame units from the data extraction unit 512, and displays an input label representing the input winner node. This is supplied to the mapping learning unit 314.

In step S362 _2, generating learning processing unit 322, for each node in the output network net Non ₁₂ in the input-output relation model M ₁₁₁₂ stored in the storage unit 511, obtains a score for output data of a frame unit from the data extraction unit 512 The node having the best score is determined as the output winner node from the nodes of the output network net ₁₂ . Furthermore, the generation learning processing unit 322 updates the output network net ₁₂ in a self-organized manner based on the output winner node for the output data in units of frames from the data extraction unit 512, and outputs an output label representing the output winner node. This is supplied to the mapping learning unit 314.

  After the processing of step S362, the process proceeds to step S363, and the mapping learning unit 314 uses the frame that has not yet been set as the frame of interest among the frames of the input data as the frame of interest, and receives the input data of the frame of interest from the learning processing unit 313. An input label that represents an input winner node for is associated with an output label that represents an output winner node for output data of a frame delayed by time T from the frame of interest from the learning processing unit 313.

  That is, in step S363, in the data buffer 331 of the mapping learning unit 314, an input label representing an input winner node for input data in units of frames from the learning processing unit 313 and an output label representing an output winner node for output data in units of frames. Is temporarily stored.

  Further, in step S363, the reading unit 332 of the mapping learning unit 314 stores the input label representing the input winner node for the input data of the frame of interest among the input labels stored in the data buffer 331 and the data buffer 331. Among the output labels, the output label representing the output winner node corresponding to the output data of the frame delayed by time T from the time of the target frame is read and associated, and the associated input label and output label are set. The label set is supplied to the connection weight update unit 333.

Then, the process proceeds from step S363 to step S364, where the connection weight update unit 333 of the mapping learning unit 314 is based on the label set supplied from the reading unit 332, and the input / output relationship model M ₁₁₁₂ stored in the storage unit 511 (FIG. 42). ), The connection relationship between the node of the input network net _{11 and} the node of the output network net ₁₂ is updated.

In other words, the connection weight updating unit 333 calculates the relationship between each input node of the input network net ₁₁ in the input / output relationship model M ₁₁₁₂ stored in the storage unit 511 and the output winner node of the output network net ₁₂ represented by the output label of the label set. The connection weight is updated around the input winner node of the input network net ₁₁ represented by the input label of the label set, and the connection weight of each output node of the output network net ₁₂ and the input winner node of the input network net ₁₁ is Update the output network net _{12 around} the output winner node.

According to the learning process of the input-output relationship model M ₁₁₁₂ as described above, binding relationship between the input network net Non ₁₁ input-output relationship model M ₁₁₁₂ (FIG. 42) has an output network net Non ₂ is updated as follows The

That is, the input winner node is the winner node for the input data of the frame of interest, and the output winner node is the winner node for the output data of the frame delayed by the input / output time difference T from the frame of interest. According to the update of the connection weight, in the input / output relationship model M ₁₁₁₂ stored in the storage unit 511, the input winner node of the input network net ₁₁ with respect to the input data of the frame of interest and the frame delayed by the input / output time difference T from the frame of interest The connection weight with the output winner node of the output network net ₂ for the output data is updated so as to be stronger.

  In the learning process of FIG. 45, the process from step S362 to S364 may be repeated with the first frame to the last frame of the input data as the frame of interest, or the first frame of the input data. The process from step S362 is repeated with the frame from the first frame to the end frame as the target frame in sequence, and then the processes from step S363 and S364 are performed again sequentially from the first frame to the end frame of the input data as the target frame. May be repeated.

Next, a process of cognitive behavior performed by the robot of FIG. 41, that is, a process of generating time-series data (recognition generating process) using the input / output relationship model _M1112 will be described with reference to the flowchart of FIG.

  For example, as described in FIG. 45, when the robot learns the task of rolling the ball in front of the eyes to the left and right, and then puts the ball in front of the robot (and if necessary, the ball is rolled. ), The state of the ball is detected by the sensor 302, and time-series sensor data representing the state is supplied to the data extraction unit 315 of the data processing device 501.

  Furthermore, motor data is also supplied from the motor drive unit 303 to the data extraction unit 315 of the data processing device 501.

  In step S371, the data extraction unit 315 uses the sensor motor data, which is a time series of vectors having the sensor data from the sensor 302 and the motor data from the motor driving unit 303 as components, as input data. The input data for each frame is extracted from the series, supplied to the recognition unit 316, and the process proceeds to step S372.

In the recognition unit 316, in step S372, the score calculation unit 341 sequentially uses the input data in units of frames from the data extraction unit 315 as the input data of the frame of interest, and the storage unit 511 for the input data of the frame of interest. The score of each node of the input network net ₁₁ constituting the input / output relation model M ₁₁₁₂ stored in is calculated and supplied to the winner node determination unit 342.

Furthermore, in step S372, the winner node determination unit 342 has the highest score from the score calculation unit 341 among the nodes of the input network net ₁₁ constituting the input / output relationship model M ₁₁₁₂ stored in the storage unit 511. To the input winner node, the input label representing the input winner node is supplied to the mapping unit 317, and the process proceeds to step S373.

In step S373, the generation node determination unit 351 of the mapping unit 317 includes the input label from the winner node determination unit 342 among the nodes of the output network net ₁₂ that constitute the input / output relationship model M ₁₁₁₂ stored in the storage unit 511. The node having the strongest connection weight with the node represented by (input winner node) is determined as the generation node, the output label indicating the generation node is supplied to the generation unit 318, and the process proceeds to step S374.

In step S374, the time series generation unit 361 of the generation unit 318 outputs the output label from the generation node determination unit 351 among the nodes of the output network net ₁₂ that constitute the input / output relationship model M ₁₁₁₂ stored in the storage unit 511. Based on the time-series pattern model 21 (FIG. 7) of the generation node represented by, for example, the time-series data of the frame length is generated and output as the output data (estimated value) for the input data of the frame of interest. To do.

  The time series data as the output data is sensor motor data, and motor data of the sensor motor data is supplied from the time series generation unit 361 to the motor drive unit 303, and the motor drive unit 303 generates the time series. The arm of the robot is moved by driving the motor according to the motor data from the unit 361.

  Thereafter, the processes in steps S372 to S374 are performed on the input data in units of frames supplied from the data extraction unit 315 to the recognition unit 316.

As described above, in the robot shown in FIG. 41, the input winner node for the input data in the frame unit of the input network net _{11 and} the frame unit delayed by the input / output time difference from the time of the input data in the frame unit of the output network net ₂ . Since the output data is updated so as to increase the connection weight with the output winner node, real-time characteristics such as a task of rolling the ball in front of the right and left are required, as in the case of the robot of FIG. Can learn such tasks and perform such tasks.

That is, according to the learning process of FIG. 45 as described above, in the input / output relationship model M ₁₁₁₂ stored in the storage unit 511, the input winner node of the input network net ₁₁ for the input data of the frame of interest and the input from the frame of interest The connection weight with the output winner node of the output network net ₁₂ for the output data of the frame delayed by the output time difference is updated so as to be stronger.

Therefore, in recognition generating process of FIG. 46, the input-output relationship model M _1112, the input data of a certain frame F is given, the input-output relationship model M _1112, of the nodes of the input network net Non _1, the frame F The node corresponding to the input data becomes the input winner node, and the node corresponding to the output data of the frame delayed by the input / output time difference from the frame F among the nodes of the output network net ₂ has the highest connection weight with the input winner node. The generation node is determined as a strong node, and thereby, time-series data corresponding to the output data of the frame delayed from the frame F by the input / output time difference is generated based on the generation node.

  Therefore, in the robot of FIG. 41, as sensor motor data as input data in units of frames, that is, as output data of frames delayed by an input / output time difference from input data in units of frames with respect to the external state and the state of the robot itself. Since learning is performed so that the sensor motor data of the motor, that is, the motor data corresponding to the action to be taken thereafter and the sensor data corresponding to the external state that will be detected thereafter can be obtained, It is possible to learn a task that requires real-time characteristics, such as a task of rolling a ball in front of the right and left, and to perform such a task.

Here, the update of the connection weight by the connection weight update unit 333 of the robot of FIG. 41, that is, the input winner node for the sensor motor data as the input data of the frame unit of the input network net _{1 and} the frame unit of the output network net ₂ Updating the sensor motor data as the output data in the frame unit delayed by the input / output time difference from the time of the sensor motor data so as to increase the connection weight with the output winner node, the sensor motor data in the frame unit and the frame Learning of the input / output relation model M ₁₁₁₂ is performed so as to associate the sensor motor data of the frame unit delayed by the input / output time difference from the time of the sensor motor data of the unit.

And, at the time of cognitive behavior, when sensor motor data in frame units is given to the input / output relationship model M ₁₁₁₂ in which such learning has been performed, in the input / output relationship model M ₁₁₁₂ , for the sensor motor data in frame units, A node corresponding to (estimated value) of frame-unit sensor motor data that is delayed by an input / output time difference from the time of the frame-unit sensor motor data is obtained as a generation node. From the sensor motor data, it can be said that the sensor motor data corresponding to the generation node, that is, the sensor motor data of the future frame is predicted from the time of frame F by the input / output time difference.

As described above, in the robot shown in FIG. 41, sensor motor data as input data in frame units and sensor motor data as output data in frame units delayed by an input / output time difference from the time of the input data in frame units. The input / output relation model M ₁₁₁₂ is learned so as to be associated with each other.

Therefore, at the time of the robot's cognitive behavior (output data generation), the time series expressed by any node of the input network net ₁₁ of the input / output relationship model M ₁₁₁₂ (FIG. 42) as sensor motor data that is input data. As long as sensor motor data (known sensor motor data) having a time-series pattern that matches the pattern is input, sensor motor data that is output data appropriate to the sensor motor data is generated. Can reproduce the learned task.

  Further, in the robot of FIG. 41, sensor motor data, which is a vector having sensor data and motor data as components, is adopted as both input data and output data, and the sensor motor data as input data in frame units is used. Thus, sensor motor data as frame-unit output data delayed by the input / output time difference from the time of the frame-unit input data, that is, sensor sensor data of the future (estimated value thereof) is generated.

  Therefore, in the robot of FIG. 41, based on the sensor motor data as input data in units of frames, the current external state and the current behavior state of the robot itself are considered, so to speak, in units of the frames. Generation of sensor motor data as frame-unit output data delayed by the input / output time difference from the input data time, that is, prediction of the future external state and determination of the future action by the input / output time difference. In other words, based on sensor data, taking into account only the current external state, motor data is generated in frame units that are delayed by the input / output time difference, that is, a future action is determined by the input / output time difference. 25, and the task can be performed with higher stability compared to the robots of FIGS.

  Specifically, for example, in the task of rolling the ball to the left or right, when it is difficult to determine the action only by the state of the ball, that is, by sensor data alone (for example, in the task of rolling the ball to the left or right, the ball is in a specific state Sometimes, when a case of moving the arm from the left to the right and a case of moving the arm from the right to the left are mixed), when noise is mixed in the sensor data (for example, in the task of rolling the ball left and right, the sensor 302 However, when sensor data is output by detecting not only the ball state that should be detected but also the arm state of the robot, motor data delayed by the input / output time difference is generated based on the sensor data alone. As a result, the robot's behavior may become unstable.

  On the other hand, when generating sensor motor data delayed by an input / output time difference based on both sensor motor data, that is, sensor data and motor data, the sensor motor data delayed by the input / output time difference Considering not only the current external state but also the current behavior state of the robot itself, it is possible to take stable behavior.

  Further, when generating sensor motor data delayed by an input / output time difference based on the sensor motor data, for example, the user guides with the arm with the gain removed, and thereby starts the task. It is possible to cause the robot to perform natural and natural interaction actions.

  Next, experimental results of experiments performed on the robot shown in FIG. 41 will be described.

  For the robot shown in FIG. 41, an experiment similar to that performed for the robot shown in FIG. 25 is performed. As a result, the movement may be somewhat unstable depending on the position of the arm. It was confirmed that, when the action of rolling to the left and right was started, the action of rolling the ball left and right with the arm continued following the movement of the ball even if the movement of the ball became somewhat unstable. Therefore, it was verified that the robot shown in FIG. 41 can learn a task that requires real-time performance and perform the task.

In the robot of FIG. 41, the input network net ₁₁ and the output network net ₁₂ included in the input / output relationship model M ₁₁₁₂ stored in the storage unit 511 are the time-series pattern storage network net included in the input / output relationship model in FIG. Similar to _in and net _out , the number of nodes, links, and time-series pattern models 21 included in the nodes may be the same or different.

On the other hand, in the robot of FIG. 41, the input network net ₁₁ and the output network net ₁₂ included in the input / output relationship model M ₁₁₁₂ stored in the storage unit 511 are both self-monitoring using sensor motor data in units of frames. Updated systematically.

Therefore, as the output network net Non ₁₂ and the input network net Non _11, when employing the same time series pattern storage network, the input network net Non ₁₁ and output network net Non ₁₂ shall be replaced by a single time series pattern storage network Can do.

That is, FIG. 47 shows an input network net Non ₁₁ and output network net Non _12, 1 single time series pattern storage input-output relationship model M _'1112 was replaced by a network input-output relationship model M ₁₁₁₂ has a Figure 42.

The input / output relationship model M ′ _{1112 in} FIG. 47 has one time-series pattern storage network (input network) net ₁₁ , and the two input networks net ₁₁ and outputs that the input / output relationship model M _{1112 in} FIG. 42 has. The network net ₁₂ is replaced by one time-series pattern storage network net ₁₁ .

In the input / output relationship model M ′ ₁₁₁₂ in FIG. 47, the input network net ₁₁ and the output network net _{12 included in the} input / output relationship model M _{1112 in} FIG. 42 are substituted by one time-series pattern storage network net ₁₁ . Therefore, the nodes of the time series pattern storage network net ₁₁ are connected by the connection weight (w ₁₁₁₂ ).

FIG. 48 shows a configuration example of a robot that performs task learning and cognitive behavior using the input / output relationship model M ′ ₁₁₁₂ in FIG. 47.

  In the figure, portions corresponding to those of the robot in FIG. 41 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  That is, the robot shown in FIG. 48 replaces the storage unit 511, the readout unit 332, the coupling weight update unit 333, and the generation node determination unit 351 shown in FIG. 41, respectively, with a storage unit 2511, a readout unit 2332, a coupling weight update unit 2333, A generation node determination unit 2351 is provided, and the configuration is the same as that of the robot of FIG. 41 except that the generation learning processing unit 322 of FIG. 41 is not provided.

The storage unit 2511 stores the input / output relationship model M ′ ₁₁₁₂ illustrated in FIG. 47.

The reading unit 2332 sequentially uses the frame of sensor motor data as input data or output data as the frame of interest, and the node label stored in the data buffer 331, that is, the recognition learning processing unit 321 outputs the sensor motor data in frame units. Of the node labels representing the winner nodes obtained in the self-organizing update of the time-series pattern storage network net ₁₁ of the input / output relationship model M ′ ₁₁₁₂ (FIG. 47) stored in the storage unit 2511, A node label representing a winner node for the sensor motor data of the frame of interest (hereinafter referred to as a first node label as appropriate) and a node label representing a winner node for the sensor motor data of a frame delayed by an input / output time difference from the time of the frame of interest (Hereinafter referred to as the second node label as appropriate) Mapping out look, the first label set is a set of second node label corresponding attached, supplied to the weight updating unit 2333.

Similar to the connection weight update unit 222 in FIG. 21, the connection weight update unit 2333 has the input / output relationship stored in the storage unit 2511 based on the label set supplied from the reading unit 2332 as described in FIG. The connection relationship between the nodes of the time-series pattern storage network net _{11 in} the model M ′ ₁₁₁₂ (FIG. 47) is updated by the Hebb rule or the like.

That is, the weight updating unit 2333, when each node in sequence pattern storage network net Non _11, the connection weight between the winning node of the time series pattern storage network net Non ₁₁ represented by the second node label label set, label set of The time series pattern storage network net ₁₁ represented by the first node label is updated centering on the winner node, and the time series pattern represented by each input node of the time series pattern storage network net ₁₁ and the first node label of the label set the coupling weight between the winning node of storage network net Non _11, updates about the winning node of the time series pattern storage network net Non ₁₁ represented by the second node label label set.

  The generation node determination unit 2351 determines a generation node that generates sensor motor data, which is time-series data, based on the node label representing the winner node supplied from the recognition unit 316, and a node label that represents the generation node Is supplied to the generation unit 318.

That is, the generation node determination unit 2351 receives from the winner node determination unit 342 of the recognition unit 316 a node label (first record) representing a winner node for sensor motor data in frame units among the nodes of the time-series pattern storage network net ₁₁ . Node labels) are supplied.

The generation node determination unit 2351 determines, as a generation node, a node having the strongest combination with the winner node represented by the node label from the winner node determination unit 342 among the nodes of the time-series pattern storage network net ₁₁ A node label representing the node is supplied to the generation unit 318.

Next, the learning process performed by the robot of FIG. 48, that is, the learning process of the input / output relationship model M ′ ₁₁₁₂ (FIG. 47) will be described with reference to the flowchart of FIG.

  For example, if the robot learns the task of rolling the ball in front of it to the left or right, the operator moves the arm to place the ball in front of the robot, hold the robot arm, and roll the ball to the left or right. .

  In this case, the sensor 302 detects the state of the ball that rolls left and right, and time-series sensor data representing the state is supplied to the data extraction unit 512 of the data processing device 501.

  In the motor drive unit 303, motor data corresponding to the movement of the arm being moved by the operator is generated and supplied to the data extraction unit 512 of the data processing device 501.

  In step S 1361, the data extraction unit 512 extracts sensor motor data in units of frames from the sensor motor data that is a time series of vectors having the sensor data from the sensor 302 and the motor data from the motor driving unit 303 as components. And it supplies to the recognition learning process part 321 of the learning process part 313, and progresses to step S1362.

In step S1362, the recognition learning processing unit 321 of the learning processing unit 313 uses the frame-unit sensor motor data from the data extraction unit 512 to store the input / output relationship model M ′ ₁₁₁₂ stored in the storage unit 2511 (FIG. 47). Self-organized learning of the time series pattern storage network net ₁₁

That is, in step S 1362, the recognition learning processing unit 321 performs frame unit sensor from the data extraction unit 512 for each node of the time-series pattern storage network net ₁₁ in the input / output relation model M ′ ₁₁₁₂ stored in the storage unit 2511. obtains a score for motor data, when from among the nodes of the sequence pattern storage network net Non _11, the best node score is determined to winning node. Further, the recognition learning processing unit 321 updates the time-series pattern storage network net ₁₁ in a self-organized manner based on the winner node for the sensor motor data in frame units from the data extraction unit 512, and represents a node label representing the winner node Is supplied to the mapping learning unit 314.

  After the process of step S1362, the process proceeds to step S1363, and the mapping learning unit 314 uses a frame that has not yet been set as the frame of interest among the frames of the sensor motor data as the frame of interest, and the learning processing unit 313 (the recognition learning processing unit 321). ) From the node label representing the winner node for the sensor motor data of the frame of interest, that is, the first node label and the sensor motor data of the frame delayed from the frame of interest by the input / output time difference T from the learning processing unit 313. The node label representing the winner node, that is, the second node label is associated.

  That is, in step S 1363, the node label representing the winner node for the sensor motor data in frame units from the learning processing unit 313 is temporarily stored in the data buffer 331 of the mapping learning unit 314.

  In step S 1363, the reading unit 2332 of the mapping learning unit 314 includes the first node label representing the winner node for the sensor motor data of the frame of interest among the node labels stored in the data buffer 331, and the frame of the frame of interest. A second node label representing the winner node for the sensor motor data of the frame delayed by the input / output time difference T from the time is read and associated, and a label set that is a set of the associated first and second node labels is , And supplied to the connection weight update unit 2333.

Then, the process proceeds from step S 1363 to step S 1364, and the connection weight update unit 2333 of the mapping learning unit 314 is based on the label set supplied from the reading unit 2332 and the input / output relationship model M ′ ₁₁₁₂ stored in the storage unit 2511 (FIG. 47) The connection relationship between the nodes of the time-series pattern storage network net ₁₁ is updated.

In other words, the connection weight update unit 2333 stores the time series pattern storage represented by each node of the time series pattern storage network net ₁₁ in the input / output relationship model M ′ ₁₁₁₂ stored in the storage unit 2511 and the second node label of the label set. the coupling weight between the winning node of the network net Non _11, each input node and updates about the winning node of the time series pattern storage network net Non ₁₁ represented by the first node label label set, the time series pattern storage network net Non ₁₁ when updates the connection weights between the time series pattern storage winning node of the network net Non ₁₁ representing the first node label, around the winning node of sequence pattern storage network net Non ₁₁ when represented by the second node label.

'According to the learning process of the _1112, input-output relationship model M' input-output relationship model M as described above coupled node relationships to each other in time series pattern storage network net Non ₁₁ having ₁₁₁₂ (FIG. 47) is found as follows Updated.

That is, the winner node represented by the first node label is the winner node for the sensor motor data of the frame of interest, and the winner node represented by the second node label is the sensor motor of the frame delayed by the input / output time difference T from the frame of interest. Since it is a winner node for the data, according to the update of the connection weight by the connection weight update unit 2333, the time series pattern storage network for the sensor motor data of the frame of interest in the input / output relation model M ′ ₁₁₁₂ stored in the storage unit 2511 and winning node of net Non _11, link weight between the winning node of the time series pattern storage network net Non ₁₁ to the sensor motor data frame delayed by output time difference T from the target frame is updated to become stronger.

  In the learning process of FIG. 49, the processes from step S1362 to S1364 may be repeated with the attention frame from the first frame to the last frame of the sensor motor data, or the beginning of the sensor motor data. The process from step S1362 is repeated with the frame from the first frame to the end frame as the target frame in sequence, and then the steps from the first frame to the end frame of the sensor motor data are sequentially performed as the target frame again at steps S1363 and You may make it perform the process of S1364 repeatedly.

Next, a process of cognitive behavior performed by the robot of FIG. 48, that is, a process of generating time-series data (recognition generating process) using the input / output relation model M ′ ₁₁₁₂ will be described with reference to the flowchart of FIG.

  For example, as described with reference to FIG. 49, when the robot learns the task of rolling the ball in front of the eyes to the left and right, and then puts the ball in front of the robot (and if necessary, the ball is rolled. ), The state of the ball is detected by the sensor 302, and time-series sensor data representing the state is supplied to the data extraction unit 315 of the data processing device 501.

  In step S 1371, the data extraction unit 315 inputs a time series of vectors having the sensor data from the sensor 302 and the motor data from the motor drive unit 303 as components as input data, and inputs the frame unit from the input data. Data is extracted and supplied to the recognition unit 316, and the process proceeds to step S1372.

In the recognition unit 316, in step S1372, the score calculation unit 341 sequentially uses the input data in units of frames from the data extraction unit 315 as the input data of the frame of interest, and the storage unit 2511 for the input data of the frame of interest. The score of each node of the time-series pattern storage network net ₁₁ constituting the input / output relation model M ′ ₁₁₁₂ (FIG. 47) stored in is calculated and supplied to the winner node determination unit 342.

In step S 1372, the winner node determination unit 342 determines the score from the score calculation unit 341 among the nodes of the time-series pattern storage network net ₁₁ constituting the input / output relationship model M ′ ₁₁₁₂ stored in the storage unit 2511. Is determined as the winner node, the node label representing the winner node is supplied to the mapping unit 317, and the process proceeds to step S1373.

In step S 1373, the generation node determination unit 2351 of the mapping unit 317 includes the winner node determination unit 342 among the nodes of the time-series pattern storage network net ₁₁ constituting the input / output relationship model M ′ ₁₁₁₂ stored in the storage unit 2511. The node having the strongest connection weight with the node represented by the node label (winner node) is determined as a generation node, the node label indicating the generation node is supplied to the generation unit 318, and the process proceeds to step S1374.

In step S 1374, the time series generation unit 361 of the generation unit 318 generates the generation node determination unit 2351 among the nodes of the time series pattern storage network net ₁₁ constituting the input / output relationship model M ′ ₁₁₁₂ stored in the storage unit 2511. Based on the time-series pattern model 21 (FIG. 7) of the generation node represented by the node label from, for example, the sensor motor data of the length of the frame as the output data (estimated value) for the input data of the frame of interest Generate and output.

  Motor data, which is a component of sensor motor data as output data, is supplied to the motor drive unit 303, and the motor drive unit 303 drives the motor based on the motor data from the time series generation unit 361. As a result, the robot arm is moved.

  Thereafter, the processes in steps S1372 to S1374 are performed on the input data in units of frames supplied from the data extraction unit 315 to the recognition unit 316.

As described above, in the robot of FIG. 48, the winner node for the frame-unit sensor motor data in the time-series pattern storage network net ₁₁ and the frame-unit sensor delayed by the input / output time difference from the time of the frame-unit sensor motor data. Since the update is performed so as to increase the connection weight with the winner node with respect to the motor data, for example, a real-time property such as a task of rolling the ball in front of the left and right is required as in the case of the robot of FIG. Learn tasks and be able to do such tasks.

Further, in the robot of FIG. 48, the input network net Non ₁₁ and output network net Non ₁₂ input-output relationship model M ₁₁₁₂ of Figure 42 that the robot of FIG. 41 is utilized with, and replaced by one time series pattern storage network net Non ₁₁ Since the input / output relation model M ′ ₁₁₁₂ (FIG. 47) is used, the storage capacity required to store one time-series pattern storage network (output network net ₁₂ ) can be saved.

  Next, for example, when the same sensor motor data is adopted as input data and output data as in the robot of FIG. 41 (the same applies to the robot of FIG. 48), learning is performed on the sensor motor data. As a score to be obtained at the time or at the time of cognitive behavior, for example, a weighted addition value obtained by weighted addition of a score for sensor data and a score for remaining motor data, which is a part of the component of sensor motor data (hereinafter, , Also referred to as a weighting score as appropriate).

  For the weighting score, for example, the calculation of the score for the sensor motor data using the time series pattern model 21 (FIG. 7) such as one HMM of the node is used as the sensor data and the motor data of the sensor motor data. These can be obtained by adding necessary weights.

  Alternatively, for example, a node is provided with two time series pattern models, a time series pattern model for sensor data and a time series pattern model for motor data, and a score obtained using the time series pattern model for sensor data And a score obtained using a time-series pattern model for motor data can be weighted to obtain a weighted score.

  In addition, about the weight attached | subjected to the score with respect to sensor data and the score with respect to motor data, the adjustment parameter for adjustment of a weight is introduce | transduced, for example, sensor data is set by setting an adjustment parameter according to a user's operation. It is possible to set which of the motor data and the motor data is to be weighted, and further to set the degree of weighting for each task, for example. By appropriately setting the adjustment parameters, it is possible to execute a stable task.

  Here, experimental results have been obtained in which the cognitive behavior performed by the robot differs depending on which of the sensor data and the motor data is emphasized.

  Next, the series of processes described above can be performed by dedicated hardware or by software. When a series of processing is performed by software, a program constituting the software is installed in a general-purpose computer, a microcomputer (controller), or the like.

  Therefore, FIG. 51 shows a configuration example of an embodiment of a computer in which a program for executing the series of processes described above is installed.

  The program can be recorded in advance in a hard disk 10105 or a ROM 10103 as a recording medium built in the computer.

  Alternatively, the program is temporarily stored in a removable recording medium 10111 such as a flexible disk, a CD-ROM (Compact Disc Read Only Memory), an MO (Magneto Optical) disk, a DVD (Digital Versatile Disc), a magnetic disk, or a semiconductor memory. It can be stored permanently (recorded). Such a removable recording medium 10111 can be provided as so-called package software.

  The program is installed on the computer from the removable recording medium 10111 as described above, or transferred from the download site to the computer wirelessly via a digital satellite broadcasting artificial satellite, or a LAN (Local Area Network), The program can be transferred to a computer via a network such as the Internet. The computer can receive the program transferred in this way by the communication unit 10108 and install it in the built-in hard disk 10105.

  The computer includes a CPU (Central Processing Unit) 10102. An input / output interface 10110 is connected to the CPU 10102 via a bus 10101, and the CPU 10102 is operated by an input unit 10107 including a keyboard, a mouse, a microphone, and the like by the user via the input / output interface 10110. When a command is input as a result, the program stored in a ROM (Read Only Memory) 10103 is executed accordingly. Alternatively, the CPU 10102 may be a program stored in the hard disk 10105, a program transferred from a satellite or a network, received by the communication unit 10108, installed in the hard disk 10105, or a removable recording medium 10111 installed in the drive 10109. The program read and installed in the hard disk 10105 is loaded into a RAM (Random Access Memory) 10104 and executed. Thus, the CPU 10102 performs processing according to the above-described flowchart or processing performed by the configuration of the above-described block diagram. Then, the CPU 10102 outputs the processing result from the output unit 10106 configured with an LCD (Liquid Crystal Display), a speaker, or the like, for example, via the input / output interface 10110, or from the communication unit 10108 as necessary. Transmission, and further recording on the hard disk 10105 and the like.

  In this specification, the processing steps for describing a program for causing a computer to perform various types of processing do not necessarily have to be processed in chronological order according to the order described in the flowchart, and are executed in parallel or individually. Processing to be performed (for example, parallel processing or object processing) is also included.

  Further, the program may be processed by a single computer, or may be processed in a distributed manner by a plurality of computers. Furthermore, the program may be transferred to a remote computer and executed.

  Here, the motor data includes, in addition to data for driving a motor, data for driving a device that works widely or a device that gives a stimulus. As a device that works to the outside or a device that gives a stimulus, there are a display that displays an image by emitting light, a speaker that outputs sound, etc. in addition to a motor. Therefore, motor data is used to drive a motor. In addition to data, image data corresponding to an image to be displayed on a display, audio data corresponding to sound output from a speaker, and the like are included.

  The embodiment of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention.

  Further, the present invention can be applied to a user interface (human interface) in, for example, an AV (Audio Visual) device such as a television receiver and an HD recorder, a computer, and other devices in addition to a robot.

It is a figure for demonstrating a control object, a forward model, and an inverse model. It is a figure which shows the assumption robot as a control object, and the forward model of the assumption robot. It is a figure which shows the arm of the robot as a control object, and the arm controller using the inverse model of the arm. It is a figure for demonstrating modeling using a linear system. It is a figure for demonstrating an example of supervised learning. It is a figure which shows the 1st structural example of a time series pattern storage network. It is a figure which shows the structural example of a node. It is a figure which shows the 2nd structural example of a time series pattern storage network. It is a figure which shows the 3rd structural example of a time series pattern storage network. It is a block diagram which shows the structural example of the data processor which performs the process using a time series pattern storage network. 3 is a block diagram illustrating a configuration example of a learning unit 4. FIG. 3 is a block diagram illustrating a configuration example of a learning processing unit 32. FIG. It is a figure for demonstrating the determination method which determines an update weight. It is a figure for demonstrating the update method which updates the learning data memorize | stored in the learning data storage part. It is a flowchart explaining a learning process. 3 is a block diagram illustrating a configuration example of a recognition unit 3. FIG. It is a flowchart explaining a recognition process. 3 is a block diagram illustrating a configuration example of a generation unit 6. FIG. It is a flowchart explaining a production | generation process. It is a figure which shows the structural example of an input / output relationship model. It is a block diagram which shows the structural example of the data processor which performs the process using an input / output relationship model. It is a flowchart explaining learning of an input / output relationship model. It is a figure which shows a connection weight matrix. It is a flowchart explaining the process which estimates output data or control data using an input-output relationship model. It is a block diagram which shows the 1st structural example of the robot which performs the task for which real-time property is requested | required. It is a diagram showing the input-output relationship model M _11. It is a figure for demonstrating the process of the data extraction part. FIG. 10 is a diagram for explaining processing of a reading unit 332; It is a flowchart illustrating the learning of the input-output relationship model M _11. Is a flowchart illustrating the recognition generating process of generating time-series data by using the input-output relationship model M _11. It is a figure which shows input data and output data when the time which considered time (DELTA) required for the process which determines a winner node is employ | adopted as an input / output time difference. It is a block diagram which shows the 2nd structural example of the robot which performs the task for which real time property is requested | required. It is a figure which shows the input / output relationship model _M123 . FIG. 10 is a diagram for explaining processing of a reading unit 432; 12 is a flowchart for explaining learning of an input / output relationship model _M123 . 10 is a flowchart for explaining recognition generation processing for generating time-series data using an input / output relationship model _M123 . Is a diagram showing the input-output relationship model M _'123. It is a block diagram which shows the 3rd structural example of the robot which performs the task for which real time property is requested | required. It is a flowchart illustrating the learning of the input-output relationship model M _'123. Is a flowchart illustrating the recognition generating process of generating time-series data by using the input-output relationship model M _'123. It is a block diagram which shows the 4th structural example of the robot which performs the task for which real time property is requested | required. It is a figure showing an input / output relationship model _M1112 . It is a figure for demonstrating the input data and output data which the data extraction part 512 handles. FIG. 10 is a diagram for explaining processing of a reading unit 332; 12 is a flowchart for explaining learning of an input / output relationship model _M1112 . 10 is a flowchart for explaining a recognition generation process for generating time-series data using an input / output relation model _M1112 . FIG. 10 is a diagram showing an input / output relationship model M ′ ₁₁₁₂ . It is a block diagram which shows the 5th structural example of the robot which performs the task for which real time property is requested | required. 10 is a flowchart for explaining learning of an input / output relation model M ′ ₁₁₁₂ . 12 is a flowchart for explaining a recognition generation process for generating time-series data using an input / output relation model _M′1112 . It is a block diagram which shows the structural example of one Embodiment of the computer to which this invention is applied.

Explanation of symbols

1 signal input unit, 2 feature extraction unit, 3 recognition unit, 4 learning unit, 5 storage unit, 6 generation unit, 21 time series pattern model, 22 learning data storage unit, 31 time series data storage unit, 32 learning processing unit, 41 score calculation unit, 42 winner node determination unit, 43 weight determination unit, 44 learning data update unit, 45 model learning unit, 51 score calculation unit, 52 winner node determination unit, 53 output unit, 61 generation node determination unit, 62:00 Sequence determination unit, 63 output unit, 211 storage unit, 212 learning unit, 213 recognition generation unit, 221 learning processing unit, 222 bond weight update unit, 231 score calculation unit, 232 winner node determination unit, 233 generation node determination unit, 234 Time series generation unit, 301 data processing device, 302 sensor, 303 motor drive unit, 311 storage unit, 312 data extraction unit, 313 learning processing unit, 314 mapping learning unit, 315 data extraction unit, 316 recognition unit, 317 mapping unit, 318 generation unit, 321 recognition learning processing unit, 322 generation learning processing unit, 331 data buffer, 332 reading unit, 333 coupling weight Update unit, 341 score calculation unit, 342 winner node determination unit, 351 generation node determination unit, 361 time series generation unit, 401 data processing device, 411 storage unit, 417 mapping unit, 421 recognition learning processing unit, 432 reading unit, 433 Connection weight update unit, 451 generation node determination unit, 501 data processing device, 511 storage unit, 512 data extraction unit, 1411 storage unit, 1432 read unit, 1433 connection weight update unit, 1451 generation node determination unit, 2511 storage unit, 2332 Read-out unit, 2333 Department, 2351 generation node determination unit, 10101 bus, 10102 CPU, 10103 ROM, 10104 RAM, 10105 hard disks, 10106 output, 10107 input unit, 10108 communication unit, 10109 drives, 10110 output interface, 10111 removable recording medium

Claims (16)

In a data processing device that learns the relationship between input data that is time-series data and output data that is other time-series data,
Input / output data extraction means for extracting input data in a predetermined time unit from the input data and extracting the output data in the predetermined time unit from the output data;
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data, and a second input time series pattern storage Network,
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output A node that is most suitable for the input data of the predetermined time unit among the nodes of the first input time-series pattern storage network in the input / output relation model that is coupled to the node of the time-series pattern storage network. A first recognition learning processing means for determining one input winner node and updating the first input time-series pattern storage network in a self-organized manner based on the first input winner node;
A second input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the second input time-series pattern storage network, and based on the second input winner node A second recognition learning processing means for updating the second input time-series pattern storage network in a self-organizing manner;
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. Generation learning processing means for self-organizing updating,
The first input winner node for the input data of the predetermined time unit of the first input time series pattern storage network and the input data of the predetermined time unit of the second input time series pattern storage network A process of updating so as to increase the coupling weight indicating the degree of coupling with the second input winner node for the input data in the predetermined time unit delayed by a certain time from the time;
The second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the second input time series pattern storage network, and the output time Update the sequence pattern storage network to increase the connection weight indicating the degree of connection with the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit. A data processing device comprising: a connection weight update unit that performs the following process: The input data is time-series sensor data representing an external state output by the sensor when a robot having a sensor for detecting an external state performs a predetermined action.
The data processing apparatus according to claim 1, wherein the output data is time-series motor data supplied to a motor that drives the robot when the robot makes the predetermined action. In a data processing method of a data processing device that learns a relationship between input data that is time-series data and output data that is other time-series data,
Extracting the input data in a predetermined time unit from the input data, and extracting the output data in the predetermined time unit from the output data;
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data, and a second input time series pattern storage Network,
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output A node that is most suitable for the input data of the predetermined time unit among the nodes of the first input time-series pattern storage network in the input / output relation model that is coupled to the node of the time-series pattern storage network. 1 input winner node is determined, and based on the first input winner node, the first input time series pattern storage network is updated in a self-organizing manner,
A second input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the second input time-series pattern storage network, and based on the second input winner node Updating the second input time-series pattern storage network in a self-organizing manner,
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. Updated in a self-organizing manner
The first input winner node for the input data of the predetermined time unit of the first input time series pattern storage network and the input data of the predetermined time unit of the second input time series pattern storage network A process of updating so as to increase the coupling weight indicating the degree of coupling with the second input winner node for the input data in the predetermined time unit delayed by a certain time from the time;
The second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the second input time series pattern storage network, and the output time Update the sequence pattern storage network to increase the connection weight indicating the degree of connection with the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit. A data processing method including steps for performing and steps. In a program that causes a computer to execute data processing for learning the relationship between input data that is time-series data and output data that is other time-series data,
Extracting the input data in a predetermined time unit from the input data, and extracting the output data in the predetermined time unit from the output data;
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data, and a second input time series pattern storage Network,
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output A node that is most suitable for the input data of the predetermined time unit among the nodes of the first input time-series pattern storage network in the input / output relation model that is coupled to the node of the time-series pattern storage network. 1 input winner node is determined, and based on the first input winner node, the first input time series pattern storage network is updated in a self-organizing manner,
A second input winner node which is a node most suitable for the input data of the predetermined time unit is determined from the nodes of the second input time-series pattern storage network, and based on the second input winner node Updating the second input time-series pattern storage network in a self-organizing manner,
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. Updated in a self-organizing manner
The first input winner node for the input data of the predetermined time unit of the first input time series pattern storage network and the input data of the predetermined time unit of the second input time series pattern storage network A process of updating so as to increase the coupling weight indicating the degree of coupling with the second input winner node for the input data in the predetermined time unit delayed by a certain time from the time;
The second input winner node for the input data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the second input time series pattern storage network, and the output time Update the sequence pattern storage network to increase the connection weight indicating the degree of connection with the output winner node for the output data of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit. A program that causes a computer to execute the data processing including the steps of: For input data that is time-series data, in a data processing device that generates output data that is other time-series data,
Input data extracting means for extracting input data in a predetermined time unit from the input data;
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data, and a second input time series pattern storage Network,
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output An input which is the most suitable node for the input data in the predetermined time unit from among the nodes of the first input time-series pattern storage network in the input / output relation model that is coupled to the node of the time-series pattern storage network An input winner node determining means for determining a winner node;
From the nodes of the second input time-series pattern storage network, a strongest connection node that is the strongest node with the input winner node is obtained, and from the nodes of the output time-series pattern storage network, the strongest connection node is obtained. Generation node determination means for determining a node having the strongest connection with a combination node as a generation node that generates the output data;
A data processing apparatus comprising: generation means for generating the output data based on the time-series pattern model of the generation node. The input data is time-series sensor data representing an external state output by a sensor included in the robot that detects an external state.
The data processing apparatus according to claim 5, wherein the output data is time-series motor data supplied to a motor that drives the robot. In the data processing method of the data processing apparatus for generating output data that is other time-series data with respect to input data that is time-series data,
Extracting input data in a predetermined time unit from the input data,
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data, and a second input time series pattern storage Network,
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output An input which is the most suitable node for the input data in the predetermined time unit from among the nodes of the first input time-series pattern storage network in the input / output relation model that is coupled to the node of the time-series pattern storage network Determine the winner node,
From the nodes of the second input time-series pattern storage network, a strongest connection node that is the strongest node with the input winner node is obtained, and from the nodes of the output time-series pattern storage network, the strongest connection node is obtained. A node having the strongest combination with the combination node is determined as a generation node that generates the output data;
A data processing method including a step of generating the output data based on the time-series pattern model of the generation node. In a program that causes a computer to execute data processing for generating output data that is other time-series data for input data that is time-series data,
Extracting input data in a predetermined time unit from the input data,
A first input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern that is a pattern of time series data as the input data, and a second input time series pattern storage Network,
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The node of the first input time series pattern storage network and the node of the second input time series pattern storage network are coupled, the node of the second input time series pattern storage network, and the output An input which is the most suitable node for the input data in the predetermined time unit from among the nodes of the first input time-series pattern storage network in the input / output relation model that is coupled to the node of the time-series pattern storage network Determine the winner node,
From the nodes of the second input time-series pattern storage network, a strongest connection node that is the strongest node with the input winner node is obtained, and from the nodes of the output time-series pattern storage network, the strongest connection node is obtained. A node having the strongest combination with the combination node is determined as a generation node that generates the output data;
A program that causes a computer to execute the data processing including the step of generating the output data based on the time-series pattern model of the generation node. In a data processing device that learns the relationship between input data that is time-series data and output data that is other time-series data,
Input / output data extraction means for extracting input data in a predetermined time unit from the input data and extracting the output data in the predetermined time unit from the output data;
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data;
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The nodes of the input time series pattern storage network are coupled to each other, and the nodes of the input time series pattern storage network and the nodes of the output time series pattern storage network are coupled to each other. From among the nodes of the time-series pattern storage network, determine an input winner node that is the most suitable node for the input data of the predetermined time unit, and based on the input winner node, the input time-series pattern storage network, Cognitive learning processing means for self-organizing updating;
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. Generation learning processing means for self-organizing updating,
Of the input winner nodes of the input time-series pattern storage network, the first input winner node for the input data in the predetermined time unit is delayed by a certain time from the time of the input data in the predetermined time unit Updating to increase the connection weight representing the degree of connection with the second input winner node for the input data in the predetermined time unit;
The second input winner node of the input time series pattern storage network and the output of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the output time series pattern storage network A data processing apparatus comprising: a connection weight updating unit configured to update the data so as to increase a connection weight representing a degree of connection with the output winner node for data. The input data is time-series sensor data representing an external state output by the sensor when a robot having a sensor for detecting an external state performs a predetermined action.
The data processing apparatus according to claim 9, wherein the output data is time-series motor data supplied to a motor that drives the robot when the robot makes the predetermined action. In a data processing method of a data processing device that learns a relationship between input data that is time-series data and output data that is other time-series data,
Extracting the input data in a predetermined time unit from the input data, and extracting the output data in the predetermined time unit from the output data;
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data;
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The nodes of the input time series pattern storage network are coupled to each other, and the nodes of the input time series pattern storage network and the nodes of the output time series pattern storage network are coupled to each other. From among the nodes of the time-series pattern storage network, determine an input winner node that is the most suitable node for the input data of the predetermined time unit, and based on the input winner node, the input time-series pattern storage network, Self-organizing,
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. Updated in a self-organizing manner
Of the input winner nodes of the input time-series pattern storage network, the first input winner node for the input data in the predetermined time unit is delayed by a certain time from the time of the input data in the predetermined time unit Updating to increase the connection weight representing the degree of connection with the second input winner node for the input data in the predetermined time unit;
The second input winner node of the input time series pattern storage network and the output of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the output time series pattern storage network And a process of updating the data so as to increase the coupling weight indicating the degree of coupling with the output winner node for the data. In a program that causes a computer to execute data processing for learning the relationship between input data that is time-series data and output data that is other time-series data,
Extracting the input data in a predetermined time unit from the input data, and extracting the output data in the predetermined time unit from the output data;
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data;
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The nodes of the input time series pattern storage network are coupled to each other, and the nodes of the input time series pattern storage network and the nodes of the output time series pattern storage network are coupled to each other. From among the nodes of the time-series pattern storage network, determine an input winner node that is the most suitable node for the input data of the predetermined time unit, and based on the input winner node, the input time-series pattern storage network Self-organizing,
An output winner node which is a node most suitable for the output data of the predetermined time unit is determined from the nodes of the output time series pattern storage network, and the output time series pattern storage network is determined based on the output winner node. Updated in a self-organizing manner
Of the input winner nodes of the input time-series pattern storage network, the first input winner node for the input data in the predetermined time unit is delayed by a certain time from the time of the input data in the predetermined time unit Updating to increase the connection weight representing the degree of connection with the second input winner node for the input data in the predetermined time unit;
The second input winner node of the input time series pattern storage network and the output of the predetermined time unit delayed by a predetermined time from the time of the input data of the predetermined time unit of the output time series pattern storage network A program that causes a computer to execute the data processing including the step of updating the data so as to increase a connection weight representing a degree of connection with the output winner node for data. For input data that is time-series data, in a data processing device that generates output data that is other time-series data,
Input data extracting means for extracting input data in a predetermined time unit from the input data;
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data;
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The nodes of the input time series pattern storage network are coupled to each other, and the nodes of the input time series pattern storage network and the nodes of the output time series pattern storage network are coupled to each other. An input winner node determining means for determining an input winner node which is a node most suitable for the input data of the predetermined time unit from the nodes of the time-series pattern storage network;
From the nodes of the input time series pattern storage network, find the strongest coupling node that is the strongest node with the input winner node, and from among the nodes of the output time series pattern storage network, Generating node determining means for determining a node having the strongest combination as a generating node for generating the output data;
A data processing apparatus comprising: generation means for generating the output data based on the time-series pattern model of the generation node. The input data is time-series sensor data representing an external state output by a sensor included in the robot that detects an external state.
The data processing apparatus according to claim 13, wherein the output data is time-series motor data supplied to a motor that drives the robot. In the data processing method of the data processing apparatus for generating output data that is other time-series data with respect to input data that is time-series data,
Extracting input data in a predetermined time unit from the input data,
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data;
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The nodes of the input time series pattern storage network are coupled to each other, and the nodes of the input time series pattern storage network and the nodes of the output time series pattern storage network are coupled to each other. From among the nodes of the time-series pattern storage network, determine an input winner node that is a node most suitable for the input data of the predetermined time unit,
From the nodes of the input time series pattern storage network, find the strongest coupling node that is the strongest node with the input winner node, and from among the nodes of the output time series pattern storage network, A node having the strongest coupling is determined as a generation node that generates the output data,
A data processing method including a step of generating the output data based on the time series pattern model of the generation node. In a program that causes a computer to execute data processing for generating output data that is other time-series data for input data that is time-series data,
Extracting input data in a predetermined time unit from the input data,
An input time series pattern storage network composed of a plurality of nodes having a time series pattern model representing a time series pattern which is a pattern of time series data as the input data;
An output time-series pattern storage network composed of a plurality of nodes having a time-series pattern model expressing a time-series pattern that is a pattern of time-series data as the output data,
The nodes of the input time series pattern storage network are coupled to each other, and the nodes of the input time series pattern storage network and the nodes of the output time series pattern storage network are coupled to each other. From among the nodes of the time-series pattern storage network, determine an input winner node that is a node most suitable for the input data of the predetermined time unit,
From the nodes of the input time series pattern storage network, find the strongest coupling node that is the strongest node with the input winner node, and from among the nodes of the output time series pattern storage network, A node having the strongest coupling is determined as a generation node that generates the output data,
A program that causes a computer to execute the data processing including the step of generating the output data based on the time-series pattern model of the generation node.

Images