JP2010067033A - Data processor, data processing method, and program - Google Patents

Abstract

Complex and long time series data is easily learned to generate smooth time series data.
A teacher data dividing unit 12 divides teacher data, which is time-series data, into a plurality of model learning data partially overlapping. The learning unit 14 assigns one model learning data to one learning model, and learns the learning model using the model learning data assigned to the learning model. The connectivity calculation unit 16 determines an error between the last part of the time series data generated by one learning model and the first part of the time series data generated by the other learning model after the one learning model. It is calculated as connectivity representing the appropriateness of connection with one other learning model. The present invention can be applied to learning and generation of time-series data, for example.
[Selection] Figure 1

Description

  The present invention relates to a data processing device, a data processing method, and a program, and in particular, for example, easily learning complicated time series data for a long time, and smooth time series data based on a learning result. The present invention relates to a data processing apparatus, a data processing method, and a program.

  For example, learning models for approximating time series patterns in the form of time difference equations and learning (memory) as dynamics include RNN (Recurrent Neural Network), SVR (Support Vector Regression), etc. ( For example, see Patent Documents 1 and 2.)

  For example, by performing RNN learning using time-series data, the RNN acquires (stores) (learns) a time-series pattern of time-series data used for learning.

  According to the RNN that has learned the time series pattern, the time series data of the time series pattern can be generated.

  In addition, a data generation technique for generating time-series data from an RNN included in a winner node (winner) using an RNN as a node of SOM (Self-Organizing Maps) has been proposed (for example, Patent Documents 3 and 4). See).

JP 2007-265345 JP JP 2006-318319 A JP 2007-280066 A Japanese Unexamined Patent Publication No. 2007-280067

  By the way, in RNN and the like, the storage mechanism of the time series pattern has a storage capacity limit. For this reason, complex (strong non-linearity) (multi-dimensional) and long-time time-series patterns are stored by RNNs of a certain scale, and time-series data of such time-series patterns are accurately reconstructed. It is difficult to (generate).

  That is, a large-scale RNN is required to store a complicated and long time series pattern and to generate time series data of such a time series pattern with high accuracy. However, large-scale RNN learning requires a large amount of computation and time.

  On the other hand, according to the data generation technique described above, it is possible to output long-time time-series data by sequentially determining the winner node, generating time-series data from the RNN possessed by the winner node, and sequentially outputting it. it can.

  However, in this case, when the node that becomes the winner node is switched, the time-series data becomes discontinuous before and after the switching.

  Therefore, as a method of outputting smooth time-series data, a method of suppressing switching of a node that becomes a winner node (Patent Document 3), or a portion in which time-series data is discontinuous (discontinuous portion) is smoothed. There is a method (Patent Document 4).

  However, in the method of suppressing the switching of the node that becomes the winner node, only the number of discontinuous portions is reduced, and if the node that becomes the winner node is switched, a discontinuous portion is generated.

  In addition, smoothing the discontinuous part is because the time-series data of the time-series pattern different from the time-series pattern stored in the RNN of the node that has become the winner node (the time-series pattern time not stored in the RNN). Series data) may be obtained.

  The present invention has been made in view of such a situation, and easily learns complex and long-time time-series data, and generates smooth time-series data with high accuracy based on the learning result. It is something that can be done.

  A data processing apparatus or program according to the first aspect of the present invention is a learning model for dividing time-series data into a plurality of pieces of data partially overlapping and learning a time-series pattern, wherein an internal state is A plurality of model learnings obtained by dividing the time-series data so as to be output as model learning data for use in learning of a learning model having and to assign one model learning data to one learning model Data is assigned to a plurality of learning models, and learning of a time-series pattern by the learning model is performed using the model learning data assigned to the learning model. One of the learning models is used for generating a learning model sequence used for generating time-series data. A start point model selecting means for selecting a start point model to be a start point of a Dell sequence, and an end point for selecting another one of the learning models as an end point model to be an end point of the generating model sequence For all of the plurality of learning models, the model selection means, the last part of the time series data generated by one of the learning models, and the first part of the time series data generated by the other learning model A value representing an error with respect to the data sequence is calculated as connectivity representing the appropriateness of connection of the time series pattern learned by the other learning model after the time series pattern learned by one of the learning models. A value corresponding to the connectivity obtained by performing one learning model after one learning model. As a connection cost for connecting the learning model, a generation model sequence calculation means for obtaining, as the generation model sequence, an arrangement of the learning models from the start point model to the end point model, which minimizes the cumulative value of the connection costs. , With respect to the learning model constituting the generating model sequence, the last part of the time series data generated by the learning model and the first part of the time series data generated by the learning model connected later Time series data generating means for determining an initial value of the internal state of the learning model and giving the initial value to the learning model so as to reduce an error with a data string, and generating time series data; A program for causing a computer to function as a data processing device or a data processing device.

  A data processing method according to a first aspect of the present invention is a learning model in which a data processing device divides time-series data into a plurality of pieces of partially overlapping data, and learns a time-series pattern. A plurality of models obtained by dividing the time-series data so that one model learning data is assigned to one learning model, and is output as model learning data used for learning a learning model having A plurality of learned models obtained by assigning learning data to a plurality of learning models, and learning a time-series pattern by the learning model using the model learning data assigned to the learning model. One of the learning models is a generation model that is a sequence of the learning model used for generating time-series data. Selecting as a start point model to be a start point of a Dell sequence, selecting another one of the plurality of learning models as an end point model to be an end point of the generating model sequence, and a plurality of the learning models For all, a value representing an error between the data sequence of the last part of the time series data generated by one learning model and the data sequence of the first part of the time series data generated by the other learning model Corresponding to the connectivity obtained by calculating the connectivity indicating the appropriateness of connection of the time series pattern learned by one other learning model after the time series pattern learned by one learning model As a connection cost for connecting one learning model to another learning model after the one learning model, For the learning model constituting the generation model sequence, an array of the learning models from the start point model to the end point model that minimizes the cumulative value of the continuation cost is obtained as the generation model sequence. The internal part of the learning model is reduced so as to reduce an error between the last part of the time series data generated by the time series data and the first part of the time series data generated by the learning model connected later. A data processing method including the steps of determining an initial value of a state and providing the initial value to the learning model to generate time-series data.

  In the first aspect as described above, one of the learning models after learning is selected as a starting point model, and the other learning model is used as an end point model. Selected. Further, the value corresponding to the connectivity is set as a connection cost for connecting the other learning model after the one learning model, and the accumulated value of the connection cost is minimized. The sequence of the learning models up to is obtained as the generation model sequence. For the learning model constituting the generation model sequence, the last part of the time series data generated by the learning model and the first part of the time series data generated by the learning model connected later The initial value of the internal state of the learning model is determined so as to reduce the error from the data sequence of the data, and the initial value is given to the learning model to generate time series data.

  A data processing apparatus or program according to a second aspect of the present invention is a learning model for dividing time-series data into a plurality of pieces of data partially overlapping and learning a time-series pattern, wherein an internal state is Dividing means for outputting as model learning data for use in learning of a learning model having a plurality of time series data obtained by dividing the time series data so as to assign one model learning data to one learning model Learning means for assigning the model learning data to the plurality of learning models, and learning the time-series pattern by the learning model using the model learning data assigned to the learning model; and a plurality of the learning For all models, the last part of the time series data generated by one learning model and the other one The time series pattern learned by one of the other learning models is connected after the time series pattern learned by one of the learning models for an error from the first partial data sequence of the time series data generated by the model It is a program for causing a computer to function as a data processing device or a data processing device that includes connectivity calculation means for calculating connectivity representing appropriateness.

  A data processing method according to a second aspect of the present invention is a learning model in which a data processing device divides time-series data into a plurality of pieces of overlapping data, and learns a time-series pattern. A plurality of models obtained by dividing the time-series data so that one model learning data is assigned to one learning model, and is output as model learning data used for learning a learning model having Data for learning is assigned to a plurality of the learning models, and learning of a time series pattern by the learning model is performed using the data for model learning assigned to the learning model. A data sequence of the last part of the time series data generated by one learning model and a time system generated by the other learning model Connectivity representing the adequacy of connecting the time series pattern learned by one of the other learning models after the time series pattern learned by one of the learning models of the error from the data sequence of the first part of the data Is a data processing method including a step of calculating as follows.

  In the second aspect as described above, the time series data is divided into a plurality of partially overlapping data, and is a learning model for learning a time series pattern, for learning a learning model having an internal state. Output as model learning data to be used. Further, a plurality of model learning data obtained by dividing the time series data so as to assign one model learning data to one learning model is assigned to the plurality of learning models, Learning of a time-series pattern by the learning model is performed using the model learning data assigned to the learning model. In addition, for all of the plurality of learning models, a data string of the last part of the time series data generated by one learning model and a data string of the first part of the time series data generated by the other learning model Is calculated as connectivity representing the adequacy of connection of the time series pattern learned by the other learning model after the time series pattern learned by one of the learning models.

  Note that the data processing device may be an independent device or an internal block constituting one device.

  The program can be provided by being transmitted via a transmission medium or by being recorded on a recording medium.

  According to the first and second aspects of the present invention, it is possible to easily learn complicated and long-time time-series data, and to generate smooth time-series data with high accuracy based on the learning result. it can.

[Overall configuration of data processing apparatus to which the present invention is applied]
FIG. 1 is a block diagram showing a configuration example of an embodiment of a data processing apparatus to which the present invention is applied.

  The data processing device, for example, generates time-series data for causing a real robot or the like to act (for example, data for driving an actuator) or time-series data for causing a virtual character or the like displayed on the display to act. learn. Furthermore, the data processing device generates time-series data for making a real robot or a virtual character act autonomously based on the learning result, and supplies the time series data to the robot. Control).

  That is, in FIG. 1, the data processing device includes a learning device 10 and a data generation device 20.

  Here, the data processing device can be configured by only the learning device 10 or the data generation device 20.

  In the data generation device 20, the learning device 10 performs data generation processing described later using information (data) obtained by performing learning processing described later. Therefore, when the data processing device is configured only from the data generation device 20, information necessary for the data generation processing is supplied to the data generation device 20 from the outside or stored inside the data generation device 20. It is necessary to keep.

  The learning device 10 is a learning model that learns a time-series pattern using time-series data (hereinafter also referred to as teacher data) prepared for learning a time-series pattern, and includes a plurality of learning models having internal states. The learning process is performed for all the plurality of learning models after learning to obtain connectivity indicating the appropriateness of connecting the time series patterns learned (stored) by any two learning models.

  That is, the learning device 10 learns, for example, teacher data that is complex, long-time time-series data, etc. as a learning process by sharing a plurality of learning models, and each of the plurality of learning models has a dynamics. A process of acquiring (storing) a time series pattern is performed.

  Furthermore, the learning device 10 performs a process of obtaining connectivity representing appropriateness (naturalness) (connectivity) of connecting time series patterns as dynamics acquired by each of a plurality of learning models as a learning process.

  Here, the dynamics represents a dynamic system that changes with time, and can be expressed by a specific function, for example. In the learning model, the temporal change feature of the time series data, that is, the time series pattern is stored as dynamics.

  The learning device 10 includes a teacher data storage unit 11, a teacher data division unit 12, a model learning data storage unit 13, a learning unit 14, a model parameter storage unit 15, a connectivity calculation unit 16, and a connectivity storage unit 17. The

  Teacher data is supplied to the teacher data storage unit 11 from the outside. The teacher data storage unit 11 stores (saves) teacher data supplied thereto.

  Here, as the teacher data, complicated and long-time time series data can be adopted. The teacher data may be, for example, simple and short time series data, or may be complex but time series data that is not so long.

  Further, in the data processing apparatus, for example, when generating time-series data for making an actual robot act autonomously in a certain environment, for example, teaching the action in an environment causing the robot to act. Time series data obtained by the user who actually moves the robot is used as teacher data.

  That is, data of physical quantities that can be sensed by the robot when the user is moving the robot (hereinafter also referred to as sensor data), and data (signals) (hereinafter referred to as the actuator) of the robot for movement. A time series of vectors having components such as action data) is used as teacher data.

  Here, a vector time series having the sensor data and action data as components as described above is also referred to as sensor motor data.

  The teacher data division unit 12 divides the time series data as the teacher data stored in the teacher data storage unit 11 into a plurality of data partially overlapping (this is also the time series data), and the learning model Is supplied to the model learning data storage unit 13 as model learning data used for learning.

  Here, the length (number of samples) of the plurality of model learning data obtained by dividing the teacher data in the teacher data dividing unit 12 may be the same or different. The same applies to the length of the overlap.

  However, in the following, for simplicity of explanation, it is assumed that the plurality of model learning data obtained by dividing the teacher data are all the same fixed length, and the overlap length is also the fixed length. I will do it.

  The model learning data storage unit 13 stores a plurality of model learning data from the teacher data dividing unit 12.

  The learning unit 14 assigns a plurality of model learning data stored in the model learning data storage unit 13 to a plurality of learning models so that one model learning data is assigned to one learning model. Further, the learning unit 14 obtains a model parameter that defines the learning model by performing learning of the time series pattern by the learning model using the model learning data assigned to the learning model. Then, the learning unit 14 supplies model parameters for each of the plurality of learning models to the model parameter storage unit 15.

  Here, the number N of the plurality of learning models to be learned by the learning unit 14 coincides with the number N of the plurality of learning model data obtained by the teacher data dividing unit 12.

  Therefore, for example, in the teacher data dividing unit 12, the teacher data is divided into data for model learning equal to or less than the number of learning models prepared in advance. Alternatively, the learning unit 14 generates the same number of learning models as the number of model learning data obtained by the teacher data dividing unit 12. Note that the substance of the learning model is a storage area such as a memory (for example, an instance in object-oriented programming).

  The model parameter storage unit 15 stores model parameters supplied from the learning unit 14.

  The connectivity calculation unit 16 uses the model learning data stored in the model learning data storage unit 13 and the model parameters stored in the model parameter storage unit 15 to perform a plurality of learnings performed by the learning unit 14. For all the models, connectivity is calculated and supplied to the connectivity storage unit 17.

  Note that connectivity represents the appropriateness of connecting a time series pattern learned by another learning model after a time series pattern learned by one learning model. The connectivity calculation unit 16 calculates the error between the last partial data sequence of the time series data generated by one learning model and the initial partial data sequence of the time series data generated by the other learning model. Calculate as

  The connectivity storage unit 17 stores the connectivity supplied from the connectivity calculation unit 16.

  The data generation device 20 is complex and corresponds to teacher data based on a plurality of learning models (model parameters) after learning obtained by the learning device 10 and connectivity calculated for the plurality of learning models. Data generation processing for generating long time, smooth time-series data is performed.

  That is, as a data generation process, the data generation device 20 uses one learning model among a plurality of learning models after learning as a starting point of a generation model sequence that is a sequence of learning models used for generating time-series data. A process of selecting as a starting point model is performed. Further, as the data generation process, the data generation apparatus 20 performs a process of selecting another one of the plurality of learning models as an end point model that is an end point of the generation model sequence.

  In addition, as the data generation process, the data generation device 20 performs a process of obtaining an arrangement of a certain learning model from the start point model to the end point model as a generation model sequence based on the connectivity.

  Further, the data generation device 20 performs a process of generating complex, long-time, smooth time-series data corresponding to the teacher data based on the generation model sequence as the data generation process.

  The data generation device 20 includes a current data supply unit 21, a target data supply unit 22, a start point model selection unit 23, an end point model selection unit 24, a generation model sequence calculation unit 25, a time series data generation unit 26, and time series data. The output unit 27 and the like are included.

  The current data supply unit 21 supplies current data, which is time series data, to the start point model selection unit 23 and the time series data generation unit 26.

  Here, a robot or the like controlled by the data processing apparatus provides a time series of vectors similar to that constituting the teacher data to the data processing apparatus as observable data. The current data is, for example, a plurality of consecutive samples including samples (vectors) of the current time among observable sensor motor data provided by a robot or the like controlled by the data processing device.

  Note that the number of samples constituting the current data is, for example, smaller than the number of samples constituting the model learning data.

  For example, the current data supply unit 21 extracts current data from observable sensor motor data provided by a robot or the like controlled by the data processing apparatus, and supplies the current data to the start point model selection unit 23 and the time-series data generation unit 26. To do.

  The target data supply unit 21 supplies target data that is time-series data to the end point model selection unit 24.

  Here, the target data is the same data (same dimension) as the current data, and is provided to the target data supply unit 21 from the outside such as a user, for example.

  For example, in the data generation device 20, time-series data for moving the robot from a current position where the robot controlled by the data processing device is located to a position designated from the outside such as a user (hereinafter also referred to as a target position). When generating certain sensor motor data, the sensor motor data (several samples of sensor motor data) obtained by the robot at the current position becomes the current data, and the sensor motor data that would be obtained at the target position is the target data. It becomes.

  The starting point model selection unit 23 is based on the current data from the current data supply unit 21, and among the plurality of learning models whose model parameters are stored in the model parameter storage unit 15, that is, among the learning models after learning, 1 One learning model is selected as the starting point model. Furthermore, the start point model selection unit 23 supplies a start point model ID (Identification) for specifying the start point model to the generation model sequence calculation unit 25.

  The end point model selection unit 24 is based on the target data from the target data supply unit 22, and among the plurality of learning models whose model parameters are stored in the model parameter storage unit 15, that is, among the learning models after learning, 1 One learning model is selected as the end point model. Further, the end point model selection unit 24 supplies an end point model ID for specifying the end point model to the generation model sequence calculation unit 25.

  Here, the starting point model is a learning model that is the starting point of a generating model sequence that is a sequence of learning models used for generating time-series data, and the end point model is a learning model that is an end point of a generating model sequence. is there.

  The start point model is used to generate the first part of time series data (hereinafter also referred to as generation time series data) generated by the time series data generation unit 26, and the end point model is Used to generate the last part of the series data.

  The generation model sequence calculation unit 25 has a plurality of learning models from the start point model specified by the start point model ID from the start point model selection unit 23 to the end point model specified by the end point model ID from the end point model selection unit 24. Is obtained as a generation model sequence.

  That is, the generation model sequence calculation unit 25 accumulates the connection cost using the value corresponding to the connectivity stored in the connectivity storage unit 17 as the connection cost for connecting one learning model to another learning model. The sequence of learning models from the start point model to the end point model that minimizes the value is obtained as a generation model sequence.

  The generation model sequence calculation unit 25 supplies the generation model sequence to the time series data generation unit 26.

  The time-series data generation unit 26 gives the current data from the current data supply unit 21 to the learning model (starting point model) constituting the generation model sequence from the generation model sequence calculation unit 25, thereby generating the generation model sequence. The time series data is generated in each learning model that constitutes.

  Further, the time series data generation unit 26 connects the time series data generated by each learning model constituting the generation model sequence (hereinafter also referred to as model generation data) in the order of the learning models as the generation model sequence. The generated time-series data is supplied to the time-series data output unit 27.

  The time series data generation unit 26 gives the current data from the current data supply unit 21 to the learning model constituting the generation model sequence from the generation model sequence calculation unit 25 before generating model generation data. In addition, with respect to the learning model that constitutes the generation model sequence, the last part of the time series data (model generation data) generated by the learning model and the time series generated by the learning model connected after (immediately after) The initial value of the internal state of the learning model is determined so as to reduce the error from the data string of the first part of the data.

  Then, the time series data generation unit 26 gives the initial value to the learning model, and generates time series data (model generation data). As a result, the generation time series data obtained by connecting the model generation data generated by each learning model constituting the generation model sequence in the order of the learning models as the generation model sequence becomes smooth time series data.

[Detailed Configuration Example of Learning Device 10]
FIG. 2 shows a more detailed configuration example of the learning device 10 of FIG.

  In FIG. 2, the teacher data dividing unit 12 divides the teacher data into a plurality of N pieces of model learning data.

  Here, the nth of the N pieces of model learning data in time series order is also referred to as model learning data #n.

  As described above, when the teacher data is divided into N pieces of model learning data # 1, # 2,..., #N in the teacher data dividing unit 12, the model learning data storage unit 13 The N pieces of model learning data # 1 to #N are stored.

The learning unit 14 includes a data allocating unit 41 and N arithmetic units 42 ₁ , 42 ₂ ,... 42 _N equal to the number of model learning data # 1 to #N.

  The data allocation unit 41 reads out N pieces of model learning data # 1 to #N stored in the model learning data storage unit 13. Further, the data allocation unit 41 allocates N model learning data # 1 to #N to N learning models # 1 so that one model learning data #n is allocated to one learning model #n. Assign to #N.

Then, the data allocation unit 41 supplies the model learning data #n to the calculation unit 42 _n that calculates the model parameter of the learning model #n.

The calculation unit 42 _n performs learning of the time series pattern by the learning model #n by using the model learning data #n assigned to the learning model #n, so that the model parameter # that defines the learning model #n is used. n is obtained (calculated) and supplied to the model parameter storage unit 15.

The model parameter storage unit 15 stores model parameters # 1 to #N defining the learning models # 1 to #N supplied from the calculation units 42 ₁ to 42 _N of the learning unit 14, respectively.

  The connectivity calculation unit 16 includes a model pair selection unit 51, a model parameter supply unit 52, two recognition generation units 53 and 54, a connectivity calculation unit 55, and the like.

  The model pair selection unit 51 selects an arbitrary two learning model sequence (permutation) from the N learning models # 1 to #N as a model pair and supplies the model pair to the model parameter supply unit 52.

  That is, the model pair selection unit 51 sequentially selects one learning model among the N learning models # 1 to #N as the attention model. Further, the model pair selection unit 51 selects one of the N learning models # 1 to #N as the subsequent model connected to the target model after the target model. select. The model pair selection unit 51 then supplies the model parameter supply unit 52 with the arrangement (permutation) of the model of interest and the subsequent model as a model pair.

  The model parameter supply unit 52 reads out the model parameters of the two learning models constituting the model pair from the model pair selection unit 51 from the model parameter storage unit 15. Further, the model parameter supply unit 52 is the first learning model (hereinafter also referred to as the previous model) in the sequence of two learning models constituting the model pair among the model parameters read from the model parameter storage unit 15. Are supplied to the recognition generation unit 53.

  In addition, the model parameter supply unit 52 sets the model parameters of the subsequent model (second learning model in the sequence of two learning models constituting the model pair) out of the model parameters read from the model parameter storage unit 15. To the recognition generation unit 54.

  The recognition generation unit 53 generates the previous model by setting the model parameter of the previous model from the model parameter supply unit 52 in the learning model (for example, an instance of the learning model as the previous model in object-oriented programming). Generate).

  In addition, the recognition generation unit 53 reads the model learning data assigned to the previous model from the model learning data storage unit 13 and gives it to the previous model, so that the model generation data that is time-series data is obtained from the previous model. Generate.

  Here, in the present embodiment, the learning model has an internal state, and an initial value of the internal state is given to the learning model when generating time-series data (model generation data). The model generation data generated from the learning model differs depending on the initial value of the internal state. The recognition generation unit 53 includes a data string (a plurality of samples) of the last part of the model generation data generated by the previous model, and the recognition generation unit 54. The initial value of the internal state given to the previous model is determined (updated) so that the error (hereinafter also referred to as a connection error) with the data string of the first part of the model generation data generated from the subsequent model is reduced.

  Then, the recognition generation unit 53 gives the initial value of the internal state when the connection error becomes small to the previous model, generates model generation data from the previous model, and supplies it to the connectivity calculation unit 55.

  The recognition generation unit 54 generates a post model by setting the model parameter of the post model from the model parameter supply unit 52 in the learning model (for example, an instance of the learning model as the post model in object-oriented programming). Generate).

  Further, the recognition generation unit 54 reads the model learning data assigned to the post model from the model learning data storage unit 13 and gives it to the post model, so that the model generation data that is time-series data is obtained from the post model. Generate.

  Here, similarly to the recognition generation unit 53, the recognition generation unit 54 also includes the first partial data string of the model generation data generated by the subsequent model and the last model generation data generated by the recognition generation unit 53 from the previous model. An initial value of an internal state to be given to the subsequent model is determined (updated) so that an error (connection error) with a partial data string is reduced.

  Then, the recognition generation unit 54 gives the initial value of the internal state when the connection error becomes small to the subsequent model, and then generates model generation data from the model and supplies it to the connectivity calculation unit 55.

  The connectivity calculation unit 55 includes a data string of the last part of the model generation data generated from the previous model from the recognition generation unit 53 and the first model generation data generated from the rear model from the recognition generation unit 54. Find the connection error with a part of the data string. Then, the connectivity calculation unit 55 supplies the connection error to the connectivity storage unit 17 as the connectivity of the rear model with respect to the previous model.

Here, the connectivity of the learning model #j for learning model #i, expressed as _{c ij (i = 1,2, ···} , N: j = 1,2, ···, N: i ≠ j).

The connectivity storage unit 17 stores N × N−N connectivity c _ij for N learning models supplied from the connectivity calculation unit 16 (the connectivity calculation unit 55).

[Description of learning model]
Next, a learning model used for learning in the learning apparatus 10 of FIG. 1 will be described.

  As a learning model, a dynamic system approximation model having an internal state among models capable of approximating a dynamic system (dynamic system approximation model) can be adopted.

  An example of a dynamic system approximation model having an internal state is RNN.

  FIG. 3 shows a configuration example of the RNN.

  Here, when data is input to a certain system (system), data output from the system with respect to the data is referred to as output data, and data input to the system is referred to as input data.

  In FIG. 3, the RNN is composed of three layers: an input layer, a hidden layer (intermediate layer), and an output layer. Each of the input layer, the hidden layer, and the output layer is configured by an arbitrary number of units corresponding to neurons.

In the RNN, input data _xt is input (supplied) from the outside to an input unit that is a part of the input layer. Here, the input data x _t represents the sample time t (value).

The remaining units of the input layer other than the input unit to which the input data _xt is input are context units, and the output of some units of the output layer is fed back to the context unit as a context representing the internal state. Is done.

Here, the context of the time t which is input to the context unit of the input layer when the input data x _t at time t is input to the input unit of the input layer, referred to as c _t.

The hidden layer unit performs weighted addition using predetermined weights (weights) for the input data x _t and context c _t input to the input layer, and the function of the nonlinear function using the result of the weighted addition as an argument. An operation is performed, and the operation result is output to the output layer unit.

In the output layer unit, the same processing as the hidden layer unit is performed on the data output from the hidden layer unit. Then, as described above, the context c _{t + 1} at the next time _{t + 1} is output from some units in the output layer and fed back to the input layer. Further, from the remaining units of the output layer, for example, output data corresponding to input data x _t is output.

  In other words, RNN learning is performed by, for example, giving RNN a sample at time t of certain time-series data as input data, and adding a sample at time t + 1 of the time-series data to the true of output data. The value is given as a value, and the error of the output data with respect to the true value is reduced.

In such learning is performed RNN, as output data to the input data x _t, predicted value x ^* _{t + 1} of the input data x _{t + 1} at the next time t + 1 of the input data x _t is output The

  As described above, in the RNN, the input to the unit is weighted and added. The weight (weight) used for the weighted addition is a model parameter of the RNN. The weights as model parameters of the RNN include weights from the input unit to the hidden layer unit, weights from the context unit to the hidden layer unit, weights from the hidden layer unit to the output layer unit, and the like.

  When the above RNN is adopted as a learning model, model learning data (model learning assigned to the learning model) that is time-series data is used as the true value of the input data and output data when learning the RNN. Data).

  In the learning of the RNN, the sample at the time t of the model learning data #n (t-th sample from the beginning) is used as the input data, and when given to the RNN, the time t + 1 as the output data output by the RNN The weight for reducing the prediction error of the predicted value of the sample is obtained by, for example, the BPTT (Back-Propagation Through Time) method.

  Further, when learning the RNN, the initial value of the context (hereinafter also referred to as the initial context) is, for example, the output data for the input data (the output data output by the RNN for the input data) with respect to the true value of the output data. It is determined (updated) in a self-organized manner so as to reduce the error.

  Here, being determined in a self-organized manner means that it is determined spontaneously without any external control.

  In addition, generation of time series data (model generation data) from RNN is to give externally supplied data to RNN as input data, or to give output data output by RNN to RNN as input data Is done by.

  In the following, it is assumed that the learning model is RNN.

[Explanation of teacher data division and learning model learning]
With reference to FIG. 4, the division of the teacher data by the teacher data dividing unit 12 (FIG. 1) and learning of the learning model using the model learning data obtained by the division will be described.

  FIG. 4 shows teacher data and assignment of model learning data obtained by dividing the teacher data to a learning model.

  In FIG. 4, the teacher data is a time series of vectors having two components.

  The teacher data dividing unit 12 (FIG. 1) uses S (> L) sample model learning data in which L samples overlap in order to share the learning data with a plurality of learning models. Divide into

  In FIG. 4, the teacher data is divided into four model learning data # 1 to # 4.

  Here, in the model learning data, the L sample that overlaps the model learning data and the adjacent model learning data is also referred to as an overlap portion of the model learning data.

  In the model learning data, which is a time series of S samples, the first L sample and the last L sample are overlapped parts (however, to be exact, the first model learning divided from the teacher data) In the data for use, only the last L sample is an overlap part, and in the last model learning data, only the first L sample is an overlap part).

  The learning unit 14 (FIG. 1) converts the model learning data # 1 into the learning model # 1, the model learning data # 2, the learning model # 2, the model learning data # 3, and the learning model # 3. The model learning data # 4 is assigned to the learning model # 4.

  Then, the learning unit 14 performs learning of the time series pattern by the learning model #n by using the model learning data #n assigned to the learning model #n, thereby obtaining the dynamics of the model learning data #n. Is obtained as a function approximation model of the time evolution equation according to the learning rule of the learning model #n.

  That is, when the learning model #n is an RNN, the learning unit 14 uses the model learning data #n to determine the weight that is the model parameter of the RNN (for example, at the time t of the model learning data #n. When the sample is given to the RNN as input data, a weight for reducing the prediction error of the prediction value of the sample at time t + 1 as output data output by the RNN is obtained by, for example, the BPTT method.

  Therefore, in the learning unit 14, when attention is paid to two learning models #n and # n + 1 to which the adjacent (continuous) model learning data #n and # n + 1 are assigned, the learning model # n + 1 Learning is model learning data #n where the L sample as the first overlap part matches the L sample as the last overlap part of model learning data #n used for learning learning model #n This is done using +1.

[Method of calculating connectivity]
With reference to FIG. 5, the connectivity calculation method by the connectivity calculation unit 16 (FIG. 1) will be described.

  The connectivity calculation unit 16 obtains connectivity representing connectivity (appropriateness) between time-series patterns as dynamics stored in each of the plurality of learning models # 1 to #N by learning by the learning unit 14.

  That is, the connectivity calculation unit 16 selects, from the plurality of learning models # 1 to #N, an arrangement (permutation) of two learning models #i and #j (i ≠ j) as a model pair.

  In addition, the connectivity calculation unit 16 performs the order of the overlapping portion, which is a partial data string (a plurality of samples) of the model generation data #i and #j generated by the learning models #i and #j constituting the model pair. Repeat propagation and back propagation (forward and back propagation). Thereby, the connectivity calculation unit 16 makes it possible to connect the model generation data #i and #j generated by the learning models #i and #j, respectively, as much as possible. Also called initial context).

  Here, the overlap part of the model generation data is a part corresponding to the overlap part of the model learning data used for learning the learning model.

  That is, as described with reference to FIG. 4, learning of the learning model is performed using model learning data of S samples having overlapping portions. Therefore, when a time series of S samples is generated as model generation data from a learning model, the model generation data has a portion corresponding to an overlap portion of model learning data used for learning. The part corresponding to the overlap part of the model learning data included in the model generation data is the overlap part of the model generation data.

  After calculating the optimal initial context, the connectivity calculation unit 16 assigns the respective optimal initial contexts to the learning models #i and #j to generate model generation data #i and #j.

Then, the connectivity calculation unit 16 includes the last overlap part (last L sample) of the model generated data #i generated by the first model constituting the model pair, that is, the first learning model #i of the model pair, The learning model as the previous model is the error (connection error) with the first overlap part (first L sample) of the model generation data #j generated by the second learning model #j of the model pair, that is, the second model. It is obtained as connectivity c _ij of learning model #j as a post model for #i.

  Here, in learning of the learning model, as described with reference to FIG. 4, when attention is paid to the two learning models #n and # n + 1, learning of the learning model # n + 1 is performed as the first overlap part. The L sample is performed using the model learning data #n that matches the L sample as the last overlap portion of the model learning data #n used for learning the learning model #n.

  In this case, the L sample as the last overlap part of model generation data #n generated by learning model #n and the first overlap part of model generation data # n + 1 generated by learning model # n + 1 Ideally, the L samples match, and the error (connection error) is zero.

  On the other hand, model learning data #n ′ (n ′ ≠ n, n ′) in which the first overlap portion does not match the last overlap portion of model learning data #n used for learning learning model #n ≠ n-1, n ′ ≠ n + 1), the first overlap part of the model generation data #n ′ generated by the learning model #n ′ generated by learning is the model generated by the learning model #n It does not match the last overlap part of generated data #n. The degree (degree) of the mismatch is determined by the first overlap portion of the model learning data #n ′ used for learning the learning model #n ′ and the model learning data used for learning the learning model #n. Corresponds to the extent that the last overlap part of #n does not match.

  From the above, for learning models #i and #j that make up a model pair, the last overlap part of model generation data #i generated by learning model #i and model generation data #j generated by learning model #j The degree of coincidence (disagreement) with the first overlap part of is used for learning the last overlap part of model learning data #i used for learning learning model #i and learning model #j. This corresponds to the degree to which the first overlap portion of the model learning data #j matches (does not match).

  Here, an error as a degree that the last overlap part of the model generation data #i generated by the learning model #i does not match the first overlap part of the model generation data #j generated by the learning model #j ( Connection error) as the connectivity representing the connectivity that connects the time series data of the time series pattern learned by the learning model #j immediately after the time series data of the time series pattern learned by the learning model #i In the teacher data dividing unit 12 (FIG. 1), the teacher data is divided into model learning data having an overlap portion.

  That is, the reason why the teacher data dividing unit 12 divides the teacher data into model learning data having an overlap portion is to calculate connectivity.

  With reference to FIG. 5, the calculation of connectivity by the connectivity calculation unit 16 (FIG. 1) will be further described.

  The connectivity calculation unit 16 selects a learning model #i as a previous model from one of N learning models # 1 to #N, and selects a learning model #j other than the learning model #i. Select as a post model.

  Then, the connectivity calculation unit 16 sets the first one sample of the model learning data #i assigned to the learning model #i as the first one sample of the input data of the learning model #i that is the previous model.

  Further, the connectivity calculation unit 16 sets the last one sample of the model learning data #j assigned to the learning model #j as the true value of the last one sample of the output data of the learning model #j, which is the subsequent model. To do.

  In addition, the connectivity calculation unit 16 sets random values as initial contexts of the learning model #i that is the previous model and the learning model #j that is the subsequent model.

  Then, the connectivity calculation unit 16 gives input data and initial context to the learning model #i, which is the previous model, and, for example, generates model generation data #i of S samples having the same length as the model learning data #i. Generate.

  After generating the S sample model generation data #i from the learning model #i, which is the previous model, the connectivity calculation unit 16 converts the L sample, which is the last overlap portion of the model generation data #i, into the subsequent model. Set as the first L sample of the input data of a learning model #j.

  Then, the connectivity calculation unit 16 gives the input data and the initial context to the learning model #j, which is the subsequent model, and, for example, generates the model generation data #j of S samples having the same length as the model learning data #j. Generate.

  Here, as described above, the L sample that is the last overlap part of the model generation data #i generated from the learning model #i that is the previous model is used as the first input data of the learning model #j that is the subsequent model. The generation of the model generation data #j from the learning model #j, which is the subsequent model, is the above-described forward propagation of the overlap portion.

  After generating the model generation data #j of the S sample from the learning model #j that is the rear model, the connectivity calculation unit 16 adds the last one of the output data of the rear model of the last sample of the model generation data #j. A prediction error for the true value of the sample (the last one sample of the model learning data #j assigned to the learning model #j as described above) is obtained.

  Then, the connectivity calculation unit 16 back propagates the prediction error of the last one sample of the model generation data #j to the first one sample of the model generation data #j based on the BPTT method, for example. Thus, the initial context of the learning model #j, which is the subsequent model, is updated so as to reduce the prediction error.

  After the initial context of the learning model #j is updated, the connectivity calculation unit 16 adds the input data (as described above, the last of the model generation data #i generated from the previous model learning model #i) to the learning model #j. L sample) which is the overlap portion of) and the initial context after update are given, and model generation data #j of S sample is generated.

  Furthermore, the connectivity calculation unit 16 uses the L sample that is the first overlap portion of the model generation data #j generated from the learning model #j that is the subsequent model as the last L sample of the learning model #i that is the previous model. Set as the true value of.

  Thereafter, the connectivity calculation unit 16 calculates the true value of the last L sample of the output data of the previous model (as described above) of the last L sample of the model generation data #i generated from the learning model #i that is the previous model. Then, a prediction error is obtained for the L sample that is the first overlap portion of the model generation data #j generated from the learning model #j after the initial context update.

  Then, the connectivity calculation unit 16 back-propagates the prediction error of the last L sample of the model generation data #i to the first sample of the model generation data #i based on the BPTT method, for example. Thus, the initial context of the learning model #i that is the previous model is updated so as to reduce the prediction error.

  After updating the initial context of the learning model #i, the connectivity calculation unit 16 adds the input data (as described above, the model learning data #i assigned to the learning model #i, which is the previous model), to the learning model #i. The first sample) and the updated initial context are given, and model generation data #i of S samples is generated.

  Here, as described above, the L sample that is the first overlap part of the model generation data #j generated from the learning model #j that is the subsequent model is used as the last of the output data of the learning model #i that is the previous model. Update the initial context of learning model #i so that the prediction error of the last L sample of model generation data #i for that true value is small, and set the model generation data # Generating i is the above-described back propagation of the overlap portion.

  The connectivity calculation unit 16 generates the S sample model generation data #i from the learning model #i that is the previous model, and then uses the L model that is the last overlap portion of the model generation data #i as the subsequent model. This is set as the first L sample of input data of a learning model #j, and the same processing is repeated thereafter.

  The connectivity calculating unit 16 then predicts the prediction error of the last L sample of the model generation data #i generated from the learning model #i and the last one sample of the model generation data #j generated from the learning model #j. When the prediction error converges, for example, the initial context obtained at that time is set as the optimum initial context.

  Furthermore, the connectivity calculation unit 16 gives the optimal initial context to each of the learning models #i and #j, and generates model generation data #i and #j.

Then, the connectivity calculation unit 16 converts the error (connection error) between the last L sample of the model generation data #i and the first L sample of the model generation data #j to the connectivity of the learning model #j with respect to the learning model #i. c _Calculate as _ij .

As described above, the connectivity calculation unit 16 reduces the connection error between the last L sample of the model generation data #i and the first L sample of the model generation data #j so as to reduce the connection model #i and #j. Each initial context is determined. The connection error between model generation data #i and #j generated by giving the initial context (optimal initial context) to learning models #i and #j is the connectivity of learning model #j to learning model #i. It is calculated as c _ij .

Note that connectivity c _ij , that is, the L overlap of the last overlap part of model generation data #i generated by learning model #i and the first overlap part of model generation data #j generated by learning model #j The connection error with the L sample is calculated, for example, according to Equation (1).

c _ij = Σ | y _j (t) -y _i (t + T) |
... (1)

Here, in Equation (1), Σ represents the total sum when the variable t is changed to an integer of 1 to L. Y _j (t) represents a sample at time t of model generation data #j generated from learning model #j (t sample from the top of model generation data #j), and y _i (t + T) Represents a sample at time t + T of model generation data #i generated from learning model #i.

In Equation (1), T is a value that satisfies Equation ST = L. In this case, the L sample of the last overlap part of the model generation data #i is y _i (1 + T), y _i (2 + T), ..., y _i (L + T) (= y _i (s)).

Further, the smaller the value of the connectivity c _ij in the equation (1), the model generation data #j generated from the learning model #j immediately follows the model generation data #i generated from the learning model #i ( Connecting) is more appropriate (natural).

[Operation of Learning Device 10]
The process (learning process) of the learning device 10 (FIG. 1) will be described with reference to FIG.

  In the learning process, in step S11, the teacher data storage unit 11 supplies the teacher data to the teacher data dividing unit 12, and the process proceeds to step S12.

  In step S12, the teacher data dividing unit 12 converts the time-series data from the teacher data storage unit 11 into N pieces of model learning data # 1 having an overlap portion of L samples as described with reference to FIG. Or split into #N. Further, the teacher data dividing unit 12 supplies and stores the N model learning data # 1 to #N to the model learning data storage unit 13, and the process proceeds from step S12 to step S13.

  In step S13, the learning unit 14 learns a learning model using the model learning data.

  That is, the learning unit 14 uses the model learning data # 1 to #N stored in the model learning data storage unit 13 so as to assign one model learning data #n to one learning model #n. Assign to learning models # 1 to #N. Further, the learning unit 14 performs learning of the time series pattern by the learning model #n using the model learning data #n assigned to the learning model #n, so that the model parameter that defines the learning model #n is used. Ask for #n. Then, the learning unit 14 supplies the model parameters # 1 to #N of the learning models # 1 to #N to the model parameter storage unit 15, respectively.

  Thereafter, the process proceeds from step S13 to step S14, the model parameter storage unit 15 stores the model parameters # 1 to #N supplied from the learning unit 14, and the process proceeds to step S15.

In step S 15, the connectivity calculation unit 16 uses the model learning data # 1 to #N stored in the model learning data storage unit 13 and the model parameters # 1 to #N stored in the model parameter storage unit 15. The connectivity calculation process for calculating the connectivity c _ij is performed for all the learning models # 1 to #N that have been used and learned by the learning unit 14, and the learning process ends.

[Explanation of connectivity calculation processing]
With reference to FIG. 7 thru | or FIG. 9, the connectivity calculation process performed by FIG.6 S15 is demonstrated.

  FIG. 7 is a flowchart for explaining connectivity calculation processing.

  In the connectivity calculation process, in step S21, the connectivity calculation unit 16 (FIG. 1) selects two learning models #i that are permutations not yet selected as model pairs from the N learning models # 1 to #N. The sequence of #j is selected, and the process proceeds to step S22.

  That is, the connectivity calculation unit 16 selects a learning model #i that is a previous model of the model pair from one learning model among the N learning models # 1 to #N, and other than the learning model #i. Learning model #j is selected as the model after the model pair.

  In step S22, the connectivity calculation unit 16 receives, from the model learning data storage unit 13 (FIG. 1), model learning data assigned to each of the previous model and the rear model, which are two learning models constituting the model pair. The reading and processing proceeds to step S23.

  In step S23, the connectivity calculation unit 16 reads out the model parameters of the previous model and the rear model constituting the model pair from the model parameter storage unit 15 (FIG. 1), and the process proceeds to step S24.

  In step S24, the connectivity calculation unit 16 sets the first one sample of the model learning data assigned to the previous model as the first one sample of the input data of the previous model, and the process proceeds to step S25.

  In step S25, the connectivity calculation unit 16 generates the previous model by setting the model parameters of the previous model in the learning model (for example, generates an instance of the learning model as the previous model in object-oriented programming). The process proceeds to step S26.

  In step S26, the connectivity calculation unit 16 sets the last one sample of the model learning data assigned to the subsequent model as the true value of the last one sample of the output data of the subsequent model, and the process proceeds to step S27. move on.

  In step S27, the connectivity calculation unit 16 sets the model parameters of the post model in the learning model, thereby generating the post model (for example, generating an instance of the learning model as the post model in object-oriented programming). The process proceeds to step S28.

  In step S28, the connectivity calculation unit 16 sets a random value as the initial context of each of the previous model and the subsequent model, and the process proceeds to step S31 in FIG.

  That is, FIG. 8 is a flowchart following FIG.

  In step S31, the connectivity calculation unit 16 gives the input data set in step S24 (FIG. 7) and the initial context (in this case, the initial context set in step S28) to the previous model to generate a model. Data is generated, and the process proceeds to step S32.

  In step S32, the connectivity calculation unit 16 sets the L sample that is the last overlap part of the model generation data generated from the previous model as the first L sample of the input data of the subsequent model, and the processing is performed in step S33. Proceed to

  In step S33, the connectivity calculation unit 16 gives the subsequent model the input data set in step S32 (L sample which is the last overlap portion of the model generation data generated from the previous model) and the initial context. Then, model generation data is generated, and the process proceeds to step S34.

  In step S33, the initial context given to the subsequent model is the updated initial context in step S35 when the process of step S35 described later has already been performed for the model pair. If the above process has not been performed yet, it is the initial context set in step S28 (FIG. 7).

  In step S34, the connectivity calculation unit 16 obtains a prediction error for the true value set in step S26 (FIG. 7) of the last sample of the model generation data generated from the subsequent model, and the process proceeds to step S35. move on.

  In step S35, the connectivity calculation unit 16 back-propagates the prediction error obtained in step S34 to the first one sample of the model generation data generated from the subsequent model based on the BPTT method. The initial context of the subsequent model is updated so as to decrease, and the process proceeds to step S36.

  In step S36, the connectivity calculating unit 16 generates the model generation data by giving the input data set in step S32 and the initial context updated in step S35 to the subsequent model, and the processing is performed in step S37. Proceed to

  In step S37, the connectivity calculation unit 16 sets the L sample of the first overlap portion of the model generation data generated from the subsequent model as the true value of the last L sample of the previous model, and the process proceeds to step S38. move on.

  In step S38, the connectivity calculation unit 16 calculates the true value set in step S37 of the last L sample of the model generation data generated from the previous model (model generation data generated from the subsequent model after updating the initial context). And the process proceeds to step S39.

  In step S39, the connectivity calculation unit 16 propagates the prediction error obtained in step S38 by back-propagating to the first sample of model generation data generated from the previous model based on, for example, the BPTT method. The initial context of the previous model is updated so as to reduce the error, and the process proceeds to step S41 in FIG.

  That is, FIG. 9 is a flowchart following FIG.

  In step S41, the connectivity calculation unit 16 determines whether a condition for ending the update of the initial context of the previous model and the subsequent model (hereinafter also referred to as an update end condition) is satisfied.

  Here, as the update end condition, it is possible to adopt that the prediction error obtained in steps S34 and S38 (FIG. 8) is in a state of being converged to some extent. Specifically, as the update end condition, the update of the initial context of the previous model and the subsequent model constituting the model pair is performed a predetermined number of times, and the prediction error obtained in steps S34 and S38 is: It can be adopted that there is almost no change between the previous time and this time.

  If it is determined in step S41 that the update end condition is not satisfied, the process returns to step S31 in FIG. 8, and the connectivity calculation unit 16 inputs the previous model to the input set in step S24 (FIG. 7). Data and initial context (in this case, the initial context updated in step S39 (FIG. 8)) are given to generate model generation data, and the same processing is repeated thereafter.

  If it is determined in step S41 that the update termination condition is satisfied, the connectivity calculation unit 16 sets the current model's current initial context as the previous model's optimal initial context and the subsequent model's current initial context. The process proceeds to step S42 with the initial context as the optimal initial context of the subsequent model.

  In step S42, the connectivity calculation unit 16 updates the previous model with the input data set in step S24 (FIG. 7) and the optimal initial context of the previous model (last step S39 (FIG. 8)). (Initial context) is given to generate model generation data, and the process proceeds to step S43.

  In step S43, the connectivity calculation unit 16 sets the L sample that is the last overlap portion of the model generation data generated from the previous model as the first L sample of the input data of the subsequent model, and the processing is performed in step S44. Proceed to

  In step S44, the connectivity calculation unit 16 adds the input data set in step S43 (L sample which is the last overlap part of the model generation data generated from the previous model) to the rear model and the optimal initial of the rear model. Given the context (initial context updated in the last performed step S35) to generate model generation data, the process proceeds to step S45.

  In step S45, the connectivity calculation unit 16 connects the last L sample of the model generation data generated from the previous model in step S42 and the first L sample of the model generation data generated from the subsequent model in step S44. Is determined according to equation (1).

Then, the connectivity calculation unit 16 sets the connection error as the connectivity c _ij of the subsequent model with respect to the previous model, and the process proceeds from step S45 to step S46.

In step S46, the connectivity calculation unit 16 supplies the connectivity c _ij obtained in step S45 to the connectivity storage unit 17 for storage, and the process proceeds to step S47.

  In step S47, the connectivity calculation unit 16 determines whether connectivity has been obtained by using all the permutations of the two learning models that can be taken by the N learning models # 1 to #N as model pairs.

  If it is determined in step S47 that there is still a permutation of two learning models that are not model pairs, the process returns to step S21 in FIG. 7, and the same process is repeated thereafter.

  If it is determined in step S47 that there is no permutation of two learning models that are not model pairs, the process returns.

[Detailed Configuration Example of Data Generation Device 20]
FIG. 10 shows a more detailed configuration example of the data generation apparatus 20 of FIG.

  In FIG. 10, the teacher data is divided into a plurality of N pieces of model learning data # 1 to #N, and N pieces of model learning data # 1 to #N are used. It is assumed that learning of learning models # 1 to #N has already been completed, including calculation of connectivity.

The start point model selection unit 23 includes a current data distribution unit 61, a model parameter supply unit 62, N recognition generation units 63 ₁ to 63 _N , a start point model determination unit 64, and the like.

The current data distribution unit 61 supplies (distributes) the current data supplied from the current data supply unit 21 to the start point model selection unit 23 to all the _N recognition generation units 63 ₁ to 63 _N.

The model parameter supply unit 62 reads the model parameters # 1 to #N of the N learning models # 1 to #N from the model parameter storage unit 15. Further, the model parameter supply unit 62 supplies the model parameter #n read from the model parameter storage unit 15 to the recognition generation unit 63 _n .

The recognition generation unit 63 _n generates a learning model #n by setting the model parameter #n from the model parameter supply unit 62 in the learning model (for example, learning using the model learning data #n has been completed). Create an instance of learning model #n in object-oriented programming).

The recognition generating unit 63 _n is the current data currently supplied from the data distribution unit 61, by giving the training model #n, the learning model #n, generates a predicted value #n current data.

  For example, a random value can be adopted as the initial context given to the learning model #n in the generation of the predicted value #n of the current data from the learning model #n. In addition, as the initial context to be given to the learning model #n, for example, an initial context (optimum initial context) for reducing the predicted value #n of the current data can be obtained and the optimal initial context can be adopted.

When the recognition generation unit 63 _n generates a prediction value #n of the current data from the learning model #n, the recognition generation unit 63 _n obtains a prediction error of the prediction value #n and supplies it to the start point model determination unit 64.

The start point model determination unit 64 selects, as start point models, the top one or more learning models with small prediction errors of the prediction values # 1 to #N of the current data supplied from the recognition generation units 63 ₁ to 63 _N , respectively. The start point model ID of the start point model is supplied to the generation model sequence calculation unit 25.

The end point model selection unit 24 includes a target data distribution unit 71, a model parameter supply unit 72, N recognition generation units 73 ₁ to 73 _N , an end point model determination unit 74, and the like.

The target data distribution unit 71 supplies (distributes) the target data supplied from the target data supply unit 22 to the end point model selection unit 24 to all the _N recognition generation units 73 ₁ to 73 _N.

The model parameter supply unit 72 reads the model parameters # 1 to #N of the N learning models # 1 to #N from the model parameter storage unit 15. Further, the model parameter supply unit 72 supplies the model parameter #n read from the model parameter storage unit 15 to the recognition generation unit 73 _n .

The recognition generation unit 73 _n generates the learning model #n by setting the model parameter #n from the model parameter supply unit 72 in the learning model.

Then, the recognition generation unit 73 _n generates the predicted value #n of the target data from the learning model #n by giving the target data supplied from the target data distribution unit 71 to the learning model #n.

  In the generation of the target data prediction value #n from the learning model #n, the initial context given to the learning model #n is a random value or the same as in the generation of the current data prediction value #n. The optimal initial context can be employed.

When the recognition generation unit 73 _n generates the prediction value #n of the target data from the learning model #n, the recognition generation unit 73 _n obtains a prediction error of the prediction value #n and supplies it to the end point model determination unit 74.

Endpoint model determination unit 74, are supplied from the recognition generating unit 73 ₁ to 73 _N, the predicted values # 1 through #N learning model prediction error is small upper one or more of the target data is selected as the end point model The end point model ID of the end point model is supplied to the generation model sequence calculation unit 25.

  The generation model sequence calculation unit 25 includes a start point model ID supply unit 81, an end point model ID supply unit 82, a sequence calculation unit 83, and the like.

  The start point model ID supply unit 81 receives the start point model ID supplied to the generation model sequence calculation unit 25 from the start point model selection unit 23 (the start point model determination unit 64), and supplies it to the sequence calculation unit 83.

  The end point model ID supply unit 82 receives the end point model ID supplied from the end point model selection unit 24 (the end point model determination unit 74) to the generation model sequence calculation unit 25 and supplies it to the sequence calculation unit 83.

  The sequence calculation unit 83 includes a plurality of learning models from the start point model specified by the start point model ID from the start point model ID supply unit 81 to the end point model specified by the end point model ID from the end point model ID supply unit 82. A certain sequence is obtained as a model sequence for generation.

That is, the sequence calculation unit 83 uses the cost corresponding to connecting the learning model #j to the value corresponding to the connectivity c _ij stored in the connectivity storage unit 17 after the learning model #i (hereinafter also referred to as connection cost). ), A sequence of learning models from the start point model to the end point model that minimizes the cumulative value of the connection cost is obtained as a generation model sequence.

  Then, the sequence calculation unit 83 supplies the generation model sequence to the time series data generation unit 26.

Here, as described above, the sequence calculation unit 83 employs a value corresponding to the connectivity c _ij as the connection cost required to connect the learning model #j after the learning model #i, and calculates the connection cost. By using a general route search algorithm, a generation model sequence that is an array of learning models from the start point model to the end point model that minimizes the cumulative connection cost by considering the distance between nodes (learning models) Ask.

  As a route search algorithm for obtaining the generation model sequence, for example, the Dijkstra method or the Viterbi algorithm can be employed.

  The generation model sequence calculation unit 25 is supplied when a plurality of start point model IDs are supplied from the start point model selection unit 23, or when a plurality of end point model IDs are supplied from the end point model selection unit 24, that is, When a plurality of learning models are selected as a start point model or an end point model, a generation model sequence is calculated for all combinations of the plurality of start points and end points.

  That is, when the number of learning models selected as the start point model is represented as A and the number of learning models selected as the end point model is represented as B, the generation model sequence calculation unit 25 generates A × B pieces. A generation model sequence is calculated.

  Then, the generation model sequence calculation unit 25 uses the generation model sequence having the minimum cumulative connection cost among the A × B generation model sequences as a generation model sequence used for generating time-series data. It is determined and supplied to the time-series data generation unit 26.

The time series data generation unit 26 includes a sequence supply unit 91, a model parameter supply unit 92, N recognition generation units 93 ₁ to 93 _N , an integrated generation unit 94, and the like.

  The sequence supply unit 91 receives the generation model sequence supplied from the generation model sequence calculation unit 25 (the sequence calculation unit 83) and supplies the generated model sequence to the model parameter supply unit 92.

The model parameter supply unit 92 reads out model parameters of a learning model (hereinafter also referred to as a configuration model) constituting the generation model sequence from the sequence supply unit 91 from the model parameter storage unit 15, and recognizes and generates the generation units 93 ₁ to 93. Supply the necessary blocks of _N.

  That is, if the generating model sequence is configured by an array of K (≦ N) configuration models # 1 to #K, the model parameter supply unit 92 includes model parameters # 1 to #K of configuration models # 1 to #K. 1 to #K are read from the model parameter storage unit 15.

Furthermore, the model parameter supply unit 92, configuration model #k (k = 1,2, ···, K) the model parameters #k of recognition generating unit 93 ₁ to the recognition generating unit 93 _k of 93 _N Supply.

The recognition generation unit 93 _k generates the configuration model #k by setting the model parameter #k from the model parameter supply unit 92 in the learning model (for example, learning using the model learning data #k has been completed). Create an instance of learning model #k in object-oriented programming).

Furthermore, the recognition generation unit 93 _k generates model generation data #k from the configuration model #k, and the recognition data generation unit 93 _{k + 1 includes} the last overlap portion of the model generation data #k and the configuration model #k. The model generation data # k + 1 generated from +1, that is, the constituent model #k so as to reduce the error from the first overlap portion of the model generation data # k + 1 connected to the model generation data #k By updating the initial context, the optimal initial context is obtained in the same manner as in the connectivity calculation process (FIGS. 7 to 9).

The recognition generating unit 93 _1, the configuration model # 1, along with providing optimal initial context of the configuration model # 1, by giving the current data supplied from the current data supplying unit 21 as input data, the model generated data # 1 is generated and supplied to the integrated generation unit 94.

Recognition generating unit 93 _k of, other than the recognition generating unit 93 ₁ of the to recognition generating unit 93 to ₁ 93 _K is the configuration model #k, with providing the optimum initial context of the configuration model #k, preceding recognition generating unit 93 _k-1 generates the model generation data #k by giving the last overlap part of the model generation data # k-1 generated from the configuration model # k-1 as the first L sample of the input data. The integrated generation unit 94 is supplied.

Integrated generator 94, the to model without generating data # 1 supplied from to 93 to ₁ recognition generating unit 93 _K #K, configuration by connecting in consideration of the overlapped portion, a smooth product time-series data (generated ) And supply to the time-series data output unit 27.

[Calculation of model sequence for generation]
Next, a method for obtaining the generation model sequence based on, for example, the Viterbi algorithm in the generation model sequence calculation unit 25 (FIG. 1) will be described.

  Here, the Viterbi algorithm is a dynamic programming algorithm that gives one most plausible explanation for the observation result. For the sequence of events (states) handled by the Viterbi algorithm, the calculation of the event at time t It is assumed that it depends only on the sequence of events at time t-1. In other words, the event handled by the Viterbi algorithm is a stochastic process based on Markov property that has the property that the future behavior is determined only by the current value and is independent of the past behavior.

  The Viterbi algorithm operates assuming a state machine. That is, the modeled system has some state at an arbitrary time. Even if the number of states is enormous, it is finite and can be listed. Each state is represented as a node. Even if a plurality of sequences (routes) of states corresponding to a given state can be considered, there is one state route that is most likely. The Viterbi algorithm examines every route that reaches a certain state and selects the most likely route. Since this is applied sequentially to the sequence of states, it is not necessary to hold every route, and it is sufficient to hold only one route per state.

  Further, in the Viterbi algorithm, an increment (usually a number) is given for a transition from one state to another. This transition is determined from the event. In the Viterbi algorithm, events are generally accumulated on the route in an additive sense. In the Viterbi algorithm, the number for each state is held, and when a certain event occurs, the best state is selected in consideration of the value of the state path so far and the increment in the new transition. The increment corresponding to the event is determined depending on the transition probability from one state to another.

  When the generation model sequence is obtained based on the Viterbi algorithm, each of the learned models # 1 to #N after learning corresponds to the state (node) of the state machine in the Viterbi algorithm. Therefore, the number N of learning models # 1 to #N after learning is the total number of states of the Viterbi algorithm.

Also, it increments corresponding to the event at the time of transition from one state to another, i.e., the transition probability in the Viterbi algorithm, the connection cost, i.e., can be used connectivity c _ij. However, the transition probability and the connection cost (connectivity c _ij ) are inversely related to the increase or decrease in value. In other words, state transitions are more likely to occur as the transition probability is larger, but the connection between learning models #i and #j, corresponding to state transitions, is likely to occur as the connection cost is smaller (mechanical connection possible) High).

In the Viterbi algorithm, a path with the maximum sum of transition probabilities among all paths from a certain starting point state to a target state is adopted as the most likely path (Vitarbi path). Similarly, in the generation of the model sequence for generation, the route with the minimum connection cost, that is, the sum of the connectivity c _ij is adopted as the route with the lowest cost, and learning corresponding to the state on the route is performed. The model sequence is a generation model sequence.

  In other words, the generation model sequence calculation unit 25 obtains, as a generation model sequence, a sequence of learning models that minimizes the cumulative value of the connection costs from the start point model to the end point model.

Now, a vector whose component is the cumulative value δ _n (τ) of the connection cost for each state #n from the first time t = 1 (the time of the state corresponding to the start point model) to a certain time t = τ, cumulative value vector _{d (τ) = (δ 1} (τ), δ 2 (τ), ···, δ N (τ)) will be referred to. The generation model sequence calculation unit 25 holds a cumulative value vector d (τ) = (δ ₁ (τ), δ ₂ (τ),..., Δ _N (τ)).

In addition, the cost of state transition from state #i to state #j, that is, the learning model corresponding to state #j immediately after learning model #i corresponding to state #i (model generation data #i generated by) the concatenation cost #j (model generation data #j that but produce) are connected, represented by b _ij. Set of connection cost b _ij is the connection cost b _ij, it can be represented by a matrix of the components of the i-th row and j column.

Here, a matrix having the connection cost b _ij as a component in the i-th row and j-th column is also referred to as a connection cost matrix.

Now, the maximum value of the connectivity c _ij which that learning model #j is connected can be considered not to be unnatural immediately after learning model #i, expressed as c _max, the maximum value c _max, connectivity c _ij The threshold is used. When the connectivity c _ij is equal to or less than the threshold value c _max , the connectivity c _ij is employed as the connection cost b _ij . Moreover, connectivity c _ij is if it exceeds the threshold value c _max is a connection cost b _ij, unreachable value c _inf is employed than the threshold value c _max is a sufficiently large value.

Here, the threshold value c _max and the inaccessible value c _inf are obtained by simulation or the like.

That is, for example, using a lot of teacher data, the last overlap part of the model generation data generated from the learning model #i and the first overlap part of the model generation data generated from the learning model #j The average value of connectivity c _{ij in} the case where it is not similar (when it is unnatural that the model generation data generated from the learning model #j is connected immediately after the model generation data generated from the learning model #i) Obtained by simulation and adopted as the threshold c _max .

In addition, for example, the maximum value of the sum of the connectivity when all of the plurality of learning models are connected using a large number of teacher data is obtained, and a value larger than the maximum value (the learning model constituting the generation model sequence) (A large value that cannot be taken as the sum of the connectivity c _ij ) is adopted as the inaccessible value c _inf .

As described above, connectivity c _ij is equal to or less than the threshold value c _max is a connection cost b _ij, if adopted connectivity c _ij, where connectivity c _ij is greater than the threshold value c _max is connected cost b By adopting the inaccessible value c _inf as _ij , it is possible to clearly distinguish between a learning model that can be connected immediately after a certain learning model and a learning model that is not connected in the generation model sequence. it can.

  In determining the generation model sequence, the generation model sequence calculation unit 25 first generates the connection cost matrix as described above and initializes the accumulated value vector d (t).

Here, initialization of the accumulated value vector d (t) means that the components δ ₁ (1), δ ₂ (1),..., Δ of the accumulated value vector d (1) at time t = 1. _N Set (set) the value of (1). In the initialization of the cumulative vector d (t), the component corresponding to the learning model that is the starting point model among the components δ ₁ (1) to δ _N (1) is set to 0, and the other components are The inaccessible value c _inf is used.

  When the generation of the connection cost matrix and the initialization of the accumulated value vector d (t) are completed, the generation model sequence calculation unit 25 performs forward calculation (calculation of the forward direction (future direction)), so that each time t The accumulated value vector d (t) is obtained.

  FIG. 11 is a diagram for explaining the forward calculation by the generation model sequence calculation unit 25.

  In FIG. 11, the horizontal axis represents time t, and the vertical axis represents a learning model corresponding to the state.

The generation model sequence calculation unit 25 _updates the component δ _j (t) of the accumulated value vector d (t) to the component δ _j (t + 1) according to the equation (2), thereby obtaining the accumulated value at time t. The vector d (t) is updated to the accumulated value vector d (t + 1) at time t + 1.

δ _j (t + 1) = min _i (δ _i (t) + b _ij )
... (2)

Here, in the formula (2), min _i () represents the minimum value of the values in parentheses when the variable i, was changed to an integer from 1 to N.

According to Equation (2), the cumulative value δ _j (t + 1) of the connection cost up to the learning model #j at time t + 1 is the cumulative value δ of the connection cost up to the learning model #i at time t + 1. _i (t) and the connection cost b _ij of the learning model #j with respect to the learning model #i.

That is, according to Equation (2), at time t + 1, the time of all connections (corresponding to state transitions) from all learning models # 1 to #N at time t that reach learning model #j. A connection (hereinafter also referred to as the minimum connection) that minimizes the cumulative value of the connection cost of t + 1 up to the learning model #j is selected. Then, via the minimum connection, at time t + 1, the cumulative value of the connection cost up to learning model #j is the cumulative value of connection cost δ _j (up to learning model #j at time t + 1. t + 1).

As a result, the generation model sequence calculation unit 25 selects only the minimum connection at time t + 1 without holding all the routes to the learning model #j, so that the learning model # at time t + 1 is selected. A cumulative value δ _j (t + 1) of connection costs up to j can be obtained.

Note that the generation model sequence calculation unit 25 uses learning formulas # 2 to #N for which the cumulative value δ _j (t + 1) of the connection cost is obtained according to Equation (2). ) Information (hereinafter also referred to as sequence information).

  That is, the generation model sequence calculation unit 25 obtains information about the learning model #i at time t (hereinafter referred to as the minimum) for each of the learning models # 1 to #N, which is the minimum connection to the learning model #j at time t + 1. (Also referred to as connection source information) is stored for each time.

  After the start of the forward calculation as described above, the generation model sequence calculation unit 25 starts determining a condition for ending the forward calculation (hereinafter also referred to as a calculation end condition), and the calculation end condition is satisfied. Finally, the forward calculation ends.

  Here, in the generation model sequence calculation unit 25, the sequence of learning models from the start point model to the end point model is obtained as a generation model sequence. How many times later from the start point model the end point model is reached Is unknown. Therefore, it is difficult to know in advance the number of times that the forward calculation should be performed. Therefore, a calculation end condition is required to end the forward calculation.

The calculation end condition is that the cumulative value δ _goal (t) of the connection cost up to the end point model among the learning models # 1 to #N is less than the threshold δ _th (formula δ _goal (t) ≦ δ _th is satisfied).

Here, in the initialization of the cumulative value vector d (t), the cumulative value of the connection cost of the starting model among the cumulative values δ ₁ (1) to δ _N (1) of the connection cost is set to 0, and the starting point A cumulative value of connection costs of learning models other than the model is set to an inaccessible value c _inf .

Therefore, for example, when the first learning model (the learning model corresponding to the state at time t = 1) in the learning model sequence (arrangement) represented by the sequence information up to the end point model is not the start point model the cumulative value of the connection cost to reaching the end point model, a connection impossible value c _inf or more.

On the other hand, when the first learning model of the learning model sequence represented by the sequence information up to the end point model becomes the start point model, that is, as a sequence of learning models from the start point model to the end point model, When an appropriate sequence of learning models with a small cumulative value is obtained, the cumulative value of the connection cost up to the end point model is the cumulative value of connectivity c _ij that is sufficiently smaller than the inaccessible value c _inf. The connection impossible value c _inf is smaller.

Therefore, when the first learning model among the series of learning models represented by the sequence information up to the end point model becomes the start point model, a general value of the cumulative value of the connection cost up to the end point model (for example, By adopting a value larger than the average value) and smaller than the inaccessible value c _inf as the threshold value δ _th and determining the calculation end condition represented by the formula δ _goal (t) ≦ δ _th An appropriate sequence of learning models from the start point model to the end point model, that is, a generation model sequence can be obtained.

The threshold value δ _th is obtained by simulation or the like. Further, as the threshold value δ _th , a fixed value can be adopted, or a variable value can be adopted. As the variable value threshold δ _th , a value that increases according to the cumulative number of connection costs (time t during forward calculation according to equation (2)) can be employed.

  When the calculation end condition is satisfied after the start of the forward calculation, the generation model sequence calculation unit 25 ends the forward calculation and performs backtrack processing to obtain the generation model sequence.

  That is, as described above, the generation model sequence calculation unit 25 stores the minimum connection source information for each learning model # 1 to #N in the forward calculation for each time.

  In the backtrack process, the generation model sequence calculation unit 25 traces the minimum connection source information from the end point model to the start point model one time at a time in the direction of going back in time. Then, the generation model sequence calculation unit 25 rearranges the minimum connection source information in the reverse order of the traced order so that the order is in the order of time, and the start point represented by the minimum connection source information in the order of the time order The sequence of learning models from the model to the end point model is obtained as a generating model sequence.

  The generation model sequence represents the order of learning models used for generating time-series data by the time-series data generating unit 26. Therefore, it can be said that the generation model sequence is a plan of the order of learning models used for generating time-series data.

[Generate time-series data using a model sequence for generation]
With reference to FIG. 12, generation of time series data (generation time series data) using a generation model sequence by the time series data generation unit 26 will be described.

  FIG. 12 shows generation time-series data generated using the generation model sequence when the generation model sequence is an arrangement of four constituent models (learning models) # 1 to # 4.

  The time-series data generation unit 26 repeats the forward propagation and the reverse propagation of the overlap portion of the model generation data, similar to the case of calculating the connectivity for the configuration models # 1 to # 4 constituting the generation model sequence. As a result, the time series data generation unit 26 makes it easy to connect the model generation data #k and # k + 1 generated by the adjacent configuration models #k and # k + 1, respectively, as much as possible. 4 Find each initial context (optimal initial context).

  Then, the time series data generation unit 26 gives the optimal initial context to the configuration models # 1 to # 4, generates model generation data # 1 to # 4 from the configuration models # 1 to # 4, and the model generation data Generate time series data by connecting # 1 to # 4.

  That is, the time-series data generation unit 26 first stores the model learning data # assigned to the configuration model # 1 as the first sample of the input data of the configuration model # 1 that is the starting point model constituting the generation model sequence. Set the first sample of i.

  Further, the time series data generation unit 26 uses the model learning data assigned to the configuration model # 4 as the true value of the last one sample of the output data of the configuration model # 4 that is the end point model constituting the generation model sequence. Set the last sample of # 4.

  In addition, the time-series data generation unit 26 sets a random value as the initial context of each of the configuration models # 1 to # 4 constituting the generation model sequence.

  Then, the time-series data generation unit 26 gives the input data and the initial context to the configuration model # 1 that is the start point model, and generates model generation data # 1 of S samples.

  After generating S-sample model generation data # 1 from the configuration model # 1 that is the starting point model, the time-series data generation unit 26 immediately follows the L sample that is the last overlap portion of the model generation data # 1. Set as the first L sample of the input data of configuration model # 2.

  Then, the time-series data generation unit 26 gives input data and an initial context to the configuration model # 2, and generates model generation data # 2 of S samples.

  Thereafter, the time-series data generation unit 26 uses the L sample that is the last overlap part of the model generation data # 2 generated from the configuration model # 2 as the first L sample of the input data of the immediately subsequent configuration model # 3. Set.

  Then, the time-series data generation unit 26 gives input data and initial context to the configuration model # 3, and generates model generation data # 3 of S samples.

  Further, the L sample which is the last overlap part of the model generation data # 3 generated from the configuration model # 3 is set as the first L sample of the input data of the immediately subsequent configuration model # 4.

  Then, the time-series data generation unit 26 gives input data and initial context to the configuration model # 4, and generates model generation data # 4 of S samples.

  As described above, when generating the model generation data # 4 from the configuration model # 4 that is the end point model, the time series data generation unit 26 outputs the configuration model # 4 of the last sample of the model generation data # 4. A prediction error for the true value of the last one sample of data (the last one sample of model learning data # 4 assigned to the constituent model # 4 as described above) is obtained.

  Then, the time-series data generation unit 26 propagates the prediction error of the last one sample of the model generation data # 4 back to the first sample of the model generation data # 4 based on, for example, the BPTT method (error propagation back) ), The initial context of the constituent model # 4 that is the end point model is updated so as to reduce the prediction error.

  After updating the initial context of the configuration model # 4, the time series data generation unit 26 adds the input data (the last of the model generation data # 3 generated from the immediately previous configuration model # 3 as described above) to the configuration model # 4. L sample) that is the overlap portion of) and the initial context after the update are given, and model generation data # 4 of S sample is generated.

  Further, the time-series data generation unit 26 sets the L sample that is the first overlap portion of the model generation data # 4 generated from the configuration model # 4 as the true value of the last L sample of the immediately previous configuration model # 3. Set.

  Thereafter, the time-series data generating unit 26 calculates the true value of the last L sample of the output data of the configuration model # 3 of the last L sample of the model generation data # 3 generated from the configuration model # 3 (as described above). Then, a prediction error is obtained for the L sample which is the first overlap portion of the model generation data # 4 generated from the learning model # 4 after the initial context update.

  Then, the time series data generation unit 26 propagates the prediction error of the last L sample of the model generation data # 3 back to the first one sample of the model generation data # 3 based on, for example, the BPTT method. ), The initial context of the configuration model # 3 is updated so as to reduce the prediction error.

  Thereafter, the time-series data generation unit 26 uses the L sample that is the first overlap portion of the model generation data # 3 generated from the configuration model # 3 as the true value of the last L sample of the immediately previous configuration model # 2. Set.

  Further, the time-series data generation unit 26 calculates the true value of the last L sample of the output data of the configuration model # 2 of the last L sample of the model generation data # 2 generated from the configuration model # 2 (as described above). Then, a prediction error is obtained for the L sample that is the first overlap portion of the model generation data # 3 generated from the learning model # 3 after the initial context update.

  Then, the time series data generation unit 26 propagates the prediction error of the last L sample of the model generation data # 2 back to the first one sample of the model generation data # 2 based on, for example, the BPTT method. The initial context of the configuration model # 2 is updated so as to reduce the prediction error.

  Thereafter, the time-series data generation unit 26 sets the L sample that is the first overlap portion of the model generation data # 2 generated from the configuration model # 2 as the true value of the last L sample of the immediately previous configuration model # 1. Set.

  Furthermore, the time-series data generation unit 26 calculates the true value of the last L sample of the output data of the configuration model # 1 of the last L sample of the model generation data # 1 generated from the configuration model # 1 (as described above). Then, a prediction error is obtained for the L sample that is the first overlap portion of the model generation data # 2 generated from the learning model # 2 after the initial context update.

  Then, the time series data generation unit 26 propagates the prediction error of the last L sample of the model generation data # 1 back to the first one sample of the model generation data # 1 based on, for example, the BPTT method. The initial context of the configuration model # 1 is updated so as to reduce the prediction error.

  As described above, when the update of the initial context from the configuration model # 4 that is the end point model to the configuration model # 1 that is the start point model is completed, the time-series data generation unit 26 stores the input data ( As described above, the first sample of the model learning data # 1 assigned to the configuration model # 1 that is the starting point model) and the updated initial context are given to generate the model generation data # 1 of the S sample To do.

  Further, the time-series data generation unit 26 uses the L sample that is the last overlap part of the model generation data # 1 generated from the configuration model # 1 that is the start point model, as the first input data of the immediately subsequent configuration model # 2. The same processing is repeated thereafter.

  Then, for example, when the prediction error obtained in each of the configuration models # 1 to # 4 constituting the generation model sequence converges, the time series data generation unit 26 uses the initial context obtained at that time as the configuration model # 1. Or # 4 as the optimal initial context for each.

  Thereafter, the time-series data generation unit 26 generates the model generation data # 1 by supporting the current data as the input data and providing the optimum initial context to the configuration model # 1 that is the start point model.

  Then, the time-series data generation unit 26 uses the L sample which is the last overlap part of the model generation data # 1 generated from the configuration model # 1 as the first L sample of the input data of the immediately subsequent configuration model # 2. Set.

  Furthermore, the time-series data generation unit 26 generates model generation data # 2 by giving input data and an optimal initial context to the configuration model # 2.

  Then, the time-series data generation unit 26 uses the L sample that is the last overlap part of the model generation data # 2 generated from the configuration model # 2 as the first L sample of the input data of the immediately subsequent configuration model # 3. Set.

  Further, the time-series data generation unit 26 generates model generation data # 3 by giving input data and an optimal initial context to the configuration model # 3.

  Then, the time-series data generation unit 26 uses the L sample that is the last overlap part of the model generation data # 3 generated from the configuration model # 3 as the first L sample of the input data of the immediately subsequent configuration model # 4. Set.

  Furthermore, the time-series data generation unit 26 generates model generation data # 4 by giving input data and an optimal initial context to the configuration model # 4.

  As described above, when the model generation data # 1 to # 4 is generated by giving the optimum initial context to each of the configuration models # 1 to # 4 constituting the generation model sequence, the time-series data generation unit 26 Connects the model generation data # 1 to # 4 to generate generation time-series data.

  That is, the time-series data generation unit 26, for example, after the model generation data #k generated from the configuration model #k, the first model generation data # k + 1 generated from the immediately subsequent configuration model # k + 1. The generation time series data is generated by connecting the samples after the overlap portion (samples after the L + 1 sample from the top of the model generation data # k + 1).

[Operation of Data Generation Device 20]
With reference to FIG. 13, the process (data generation process) of the data generation apparatus 20 will be described.

  In the data generation device 20, in step S61, the current data supply unit 21, the target data supply unit 22, the start point model selection unit 23, the end point model selection unit 24, and the generation model sequence calculation unit 25 generate the generation model sequence. A calculation process for calculating is performed.

  Further, in step S61, the generation model sequence calculation unit 25 supplies the generation model sequence obtained in the generation model sequence calculation process to the time-series data generation unit 26, and the process proceeds to step S62.

  In step S 62, the time series data generation unit 26 generates generation time series data using the generation model sequence from the generation model sequence calculation unit 25 and supplies the time series data output unit 27 to the time series data output unit 27. A process is performed, and the process proceeds to step S63.

  In step S63, the time-series data output unit 27 outputs the generated time-series data from the time-series data generation unit 26 to the robot controlled by the data processing apparatus in FIG. 1, and the data generation process ends.

  The robot controlled by the data processing apparatus of FIG. 1 is driven according to action data among the components of the generated time series data (sensor motor data) from the time series data output unit 27. As a result, the robot takes a predetermined action, that is, an action appropriate for obtaining target data from a state in which data is currently obtained as sensor data sensed by the robot.

[Generation of model sequence for generation]
With reference to FIG. 14, the generation model sequence calculation process performed in step S61 of FIG. 13 will be described.

  In step S71, the current data supply unit 21 supplies the current data to the start point model selection unit 23 and the time series data generation unit 26, and the process proceeds to step S72.

  In step S72, the start point model selection unit 23 uses the current data from the current data supply unit 21 as input data, and from each of the N learning models # 1 to #N whose model parameters are stored in the model parameter storage unit 15. Then, model generation data # 1 to #N, which are predicted values of the current data, are generated (recognized and generated).

  Then, the process proceeds from step S72 to step S73, and the start point model selection unit 23 obtains a prediction error of the prediction value of the current data for each of the model generation data # 1 to #N. Furthermore, the start point model selection unit 23 selects, for example, the top one learning model having a small prediction error from among the N learning models # 1 to #N as the start point model, and the processing is performed from step S73 to step S73. Proceed to S74.

  In step S74, the target data supply unit 22 supplies the target data to the end point model selection unit 24, and the process proceeds to step S75.

  In step S75, the end point model selection unit 24 uses the target data from the target data supply unit 22 as input data, and from each of the N learning models # 1 to #N whose model parameters are stored in the model parameter storage unit 15. Then, model generation data # 1 to #N, which are predicted values of the target data, are generated (recognized and generated).

  Then, the process proceeds from step S75 to step S76, and the end point model selection unit 24 obtains the prediction error of the predicted value of the target data for each of the model generation data # 1 to #N. Further, the end point model selection unit 24 selects, for example, the top one learning model having a small prediction error among the N learning models # 1 to #N as the end point model, and the processing is performed from step S76 to step S76. Proceed to S77.

  In step S 77, the start point model selection unit 23 supplies the start point model ID of the start point model to the generation model sequence calculation unit 25. Furthermore, in step S77, the end point model selection unit 24 supplies the end point model ID of the end point model to the generation model sequence calculation unit 25, and the process proceeds from step S77 to step S78.

  In step S 78, the generation model sequence calculation unit 25 specifies the start point model based on the start point model ID from the start point model selection unit 23 and specifies the end point model based on the end point model ID from the end point model selection unit 24.

  Further, the generation model sequence calculation unit 25 obtains a certain arrangement of a plurality of learning models from the start point model to the end point model as a generation model sequence.

  That is, as described above, the generation model sequence calculation unit 25 uses the value corresponding to the connectivity stored in the connectivity storage unit 17 as the connection cost for connecting one learning model to another learning model. The sequence of learning models from the start point model to the end point model that minimizes the cumulative value of connection costs is obtained as a generation model sequence.

  Then, the generation model sequence calculation unit 25 supplies the generation model sequence to the time series data generation unit 26, and the process returns.

[Time-series data generation processing]
With reference to FIGS. 15 to 18, the time-series data generation process performed in step S62 of FIG. 13 will be described.

  FIG. 15 is a flowchart for explaining time-series data generation processing.

  In the time-series data generation process, in step S81, the time-series data generation unit 26 receives the generation model sequence supplied from the generation model sequence calculation unit 25, and the process proceeds to step S82.

  In step S82, the time-series data generation unit 26 uses the model learning data storage unit 13 (the model learning data assigned to each of the start point model and the end point model among the constituent models constituting the generation model sequence. After reading from FIG. 1), the process proceeds to step S83.

  In step S83, the time-series data generation unit 26 reads out the model parameters of each of the constituent models constituting the generation model sequence from the model parameter storage unit 15 (FIG. 1), and the process proceeds to step S84.

  In step S84, the time-series data generation unit 26 sets the first one sample of the model learning data assigned to the start point model as the first one sample of the input data of the start point model, and the process proceeds to step S85. move on.

  In step S85, the time-series data generation unit 26 sets the last one sample of the model learning data assigned to the end point model as the true value of the last one sample of the output data of the end point model. Proceed to step S86.

  In step S86, the time-series data generation unit 26 generates a configuration model that constitutes the generation model sequence by setting model parameters of the configuration model that constitutes the generation model sequence in the learning model (for example, object In the oriented programming, an instance of a learning model as a configuration model is generated), and the process proceeds to step S87.

  In step S88, the time-series data generation unit 26 sets a random value as the initial context of each of the constituent models constituting the generation model sequence, and the process proceeds to step S91 in FIG.

  That is, FIG. 16 is a flowchart following FIG.

  In step S91, the time-series data generation unit 26 selects a start point model among the constituent models constituting the generation model sequence as a target model of interest. Further, in step S91, the input model set in step S84 (FIG. 15) and the initial context (in this case, the initial context set in step S87) are given to the starting point model that is the model of interest to generate a model. Data is generated, and the process proceeds to step S92.

  In step S92, the time-series data generation unit 26 newly sets a configuration model immediately after the current model of interest (hereinafter also referred to as a model immediately after) of the configuration models constituting the generation model sequence as a model of interest. select.

  Further, the time-series data generation unit 26 uses, as the first L sample of the input data of the model of interest, the configuration model immediately before the current model of interest among the configuration models constituting the generation model sequence (hereinafter referred to as the previous model) The L sample which is the last overlap part of the model generation data generated from the above is set, and the process proceeds from step S92 to step S93.

  In step S93, the time-series data generation unit 26 uses the input data set in step S92 (L sample that is the last overlap part of the model generation data generated from the immediately preceding model) and the initial context as the target model. Given, model generation data is generated, and the process proceeds to step S94.

  Note that in steps S91 and S93, the initial context given to the model of interest is the initial context after the update in steps S102 and S106 if the processing in steps S102 and S106 (FIG. 17) described later has already been performed. If it is a context and the processing of steps S102 and S106 has not been performed yet, it is the initial context set in step S87 (FIG. 15).

  In step S94, the time-series data generation unit 26 determines whether the model of interest is an end point model. If it is determined in step S94 that the model of interest is not the end point model, the process returns to step S92, and the same process is repeated thereafter.

  If it is determined in step S94 that the model of interest is an end point model, that is, if model generation data is generated from all of the constituent models constituting the generation model sequence, the process is as shown in FIG. Proceed to step S101.

  That is, FIG. 17 is a flowchart following FIG.

  In step S101, the time-series data generation unit 26 obtains a prediction error for the true value set in step S85 (FIG. 15) of the last one sample of the model generation data generated from the end point model. Proceed to S102.

  In step S102, the time-series data generation unit 26 propagates the prediction error obtained in step S102 back to the first sample of model generation data generated from the end point model based on the BPTT method. The initial context of the end point model is updated so as to reduce the error, and the process proceeds to step S103.

  In step S103, the time series data generation unit 26 selects the end point model as the model of interest. Further, in step S103, the time-series data generation unit 26 gives the input model set in step S92 (FIG. 16) and the updated initial context in step S102 to the end point model that is the model of interest, and the model Generate generated data.

  Then, the process proceeds from step S103 to step S104, and the time-series data generation unit 26 newly selects a model immediately before the target model as the target model. Further, in step S104, the time-series data generation unit 26 sets the L sample of the first overlap portion of the model generation data generated from the immediately preceding model as the true value of the last L sample of the model of interest. The process proceeds to step S105.

  In step S105, the time series data generation unit 26 calculates the true value set in step S104 of the last L sample of the model generation data generated from the model of interest (the model generated from the model immediately after the initial context is updated). The prediction error for L sample of the first overlap portion of the generated data is obtained, and the process proceeds to step S106.

  In step S106, the time series data generation unit 26 propagates the prediction error obtained in step S105 back to the first sample of model generation data generated from the model of interest based on, for example, the BPTT method. The initial context of the model of interest is updated so as to reduce the prediction error, and the process proceeds to step S107.

  In step S107, the time-series data generation unit 26 determines whether the model of interest is a start point model. If it is determined in step S107 that the model of interest is not the start point model, the process returns to step S104, and the same process is repeated thereafter.

  If it is determined in step S107 that the model of interest is the start point model, that is, in steps S101 to S106, all of the constituent models constituting the generation model sequence from the end point model toward the start point model are displayed. When the initial context is updated, the process proceeds to step S108, and the time-series data generation unit 26 satisfies a condition (update end condition) for ending the update of the initial context of the configuration model constituting the generation model sequence. Determine whether or not.

  Here, as the update end condition in Step S108, it can be adopted that the prediction error obtained in Steps S101 and S105 is in a state of being converged to some extent. Specifically, as the update end condition, the initial context of the configuration model constituting the generation model sequence has been updated by a predetermined number of repetitions, and the prediction error obtained in steps S101 and S105 is the previous time. And this time, it can be adopted that there is almost no change.

  If it is determined in step S108 that the update end condition is not satisfied, the process returns to step S91 in FIG. 16, and the time-series data generation unit 26 is set in the start point model in step S84 (FIG. 15). The input data and the initial context (in this case, the initial context after the update in step S106) are given to generate model generation data. Thereafter, the same processing is repeated.

  If it is determined in step S108 that the update end condition is satisfied, the time-series data generation unit 26 uses the current initial context of the configuration model as the optimal initial context of the configuration model, and the process is as shown in FIG. Proceed to step S111 of FIG.

  That is, FIG. 18 is a diagram following FIG.

  In step S111, the time-series data generation unit 26 uses the current data supplied from the current data supply unit 21 (FIG. 1) as the first plurality of samples of the input data of the start point model (the same number of samples as the current data). ) And the process proceeds to step S112.

  In step S112, the time series data generation unit 26 selects the start point model as the model of interest.

  Further, in step S112, the time-series data generation unit 26 generates the S sample model generation data by giving the input model set in step S111 and the optimal initial context of the start point model to the start point model that is the model of interest. Then, the process proceeds to step S113.

  In step S113, the time series data generation unit 26 outputs the model generation data of the S samples generated in step S112 to the time series data output unit 27 (FIG. 1) as (part of) the generation time series data. The process proceeds to step S114.

  In step S114, the time-series data generation unit 26 newly selects a model immediately after the target model as the target model.

  Further, in step S114, the time-series data generation unit 26 sets the L sample that is the last overlap portion of the model generation data generated from the model immediately before the target model as the first L sample of the input data of the target model. Then, the process proceeds to step S115.

  In step S115, the time-series data generation unit 26 adds the input data set in step S114 (L sample that is the last overlap part of the model generation data generated from the immediately preceding model) to the target model and the target model. An optimal initial context is given to generate model generation data, and the process proceeds to step S116.

  In step S116, the time series data generation unit 26 generates a sample after the L + 1 sample from the model generation data generated from the target model in step S115, following the generation time series data output immediately before. It outputs to the time series data output part 27 (FIG. 1) as series data, and a process progresses to step S117.

  In step S117, the time-series data generation unit 26 determines whether the model of interest is an end point model. If it is determined in step S117 that the model of interest is not the end point model, the process returns to step S114, and the same process is repeated thereafter.

  If it is determined in step S117 that the model of interest is an end point model, that is, if model generation data is generated from all of the constituent models constituting the generation model sequence, the process returns.

  As described above, in the data processing device 10 (FIG. 1), the teacher data dividing unit 12 divides the teacher data, which is time series data, into a plurality of partially overlapping data, and has a learning model having an internal state. Is output as model learning data used for learning. Further, the learning unit 14 assigns a plurality of model learning data to a plurality of learning models so that one model learning data is assigned to one learning model, and the learning of the time series pattern by the learning model is performed. The model learning data assigned to the learning model is used. Then, the connectivity calculation unit 16 includes, for all of the plurality of learning models, an overlap part that is a data string of the last part of the time series data generated by one learning model and a time series generated by the other learning model. The error from the first overlap portion of the data is calculated as connectivity representing the appropriateness of connection of the time series pattern learned by one other learning model after the time series pattern learned by one learning model.

  On the other hand, in the data generation device 20 (FIG. 1), the start point model selection unit 23 selects one learning model as a start point model from among a plurality of learning models after learning, and the end point model selection unit 24 One other learning model is selected as the end point model. Further, the generation model sequence calculation unit 25 uses the value corresponding to the connectivity as the connection cost for connecting another learning model after one learning model, and minimizes the cumulative value of the connection cost. The sequence of learning models from to the end model is obtained as a generation model sequence. Then, the time series data generation unit 26 generates, for the learning model (configuration model) constituting the generation model sequence, the last overlapping portion of the time series data generated by the learning model and the learning model connected later. The initial value of the internal state of the learning model is determined so as to reduce the error from the first overlap portion of the time series data, and the initial value is given to the learning model to generate time series data.

  Therefore, it is possible to easily learn complicated and long-time time-series data, and to generate smooth time-series data with high accuracy based on the learning result.

  In other words, the learning device 10 performs learning by storing in a time direction the complex (non-linear, multi-dimensional) and long-time dynamics that cannot be stored by one learning model by using a plurality of learning models. In 20, a generation model sequence that is an array of such learning models after learning is calculated, and generation time-series data is generated using the learning models constituting the generation model sequence.

  In the calculation of the model sequence for generation, the dynamics stored in each learning model are as smooth as possible based on the connectivity that is the evaluation value for the connectivity between the learning models, and the path from the start point to the end point is as short as possible. As shown in FIG. 1, the learning model sequence including the inexperienced plan (the learning model sequence other than the learning model sequence generating generation time series data corresponding to all or part of the teacher data) is obtained.

  Furthermore, in generation time series data generation, the model generation data in the forward direction is transferred by forward propagation that takes over the last overlap part of the model generation data generated from the immediately preceding model as the first part of the input data of the model of interest. On the other hand, based on the model generation data generated in the forward direction, the prediction error calculated by the end point model is propagated in the reverse direction, that is, the learning model on the start point model side, thereby generating the model sequence for generation. The initial context of the configured learning model is corrected (updated). Then, by repeating this forward and reverse propagation, the generation model sequence is a sequence of learning models other than the sequence of learning models that generate generation time-series data corresponding to all or part of the teacher data. Even in such a case, the initial context is modified so that the model generation data generated from the learning model constituting the generation model sequence is smoothly connected, and smooth generation time-series data is generated (reconstructed).

  More specifically, in the data learning device 10, the teacher data that is time-series data is divided into a plurality of model learning data that partially overlap. Then, a plurality of model learning data is assigned to a plurality of learning models so that one model learning data is assigned to one learning model, and time-series pattern learning by the learning model is assigned to the learning model. The model learning data is used.

  Therefore, since time series data is shared and learned by multiple learning models (function approximation learning), the storage capacity of time series patterns is eliminated, and complicated and long time series patterns are reduced. The time (short) time-series pattern can be divided and stored. Furthermore, it is possible to accurately generate (reconstruct) time-series data of a complicated and long-time time series pattern using a learning model that stores such a short-time time-series pattern.

  That is, since the length of the time series pattern in which one learning model is in charge of learning is limited, the time series pattern can be accurately learned (stored) even if the learning model is a small-scale RNN or the like. . Furthermore, by increasing the number of learning models, it is possible to increase the overall storage capacity of a plurality of learning models, so that complex and long time series patterns can be stored without being influenced by the storage capacity of one learning model. can do.

  In addition, the learning device 10 obtains connectivity, and the data generation device 20 calculates a generation model sequence based on the connectivity. Therefore, the model learning data for which the learning model is responsible for learning is the data at which position of the teacher data. It is possible to calculate a generation model sequence as an array of learning models used for generating time-series data without depending on whether or not.

  That is, for example, in order for a mobile robot moving in a certain environment to execute a task (navigation task) that moves from the current position to the goal position that is the goal, from the experience given as teacher data, It is necessary to make a plan of the route to travel to.

  For example, when the mobile robot moves between any two points in the environment in which the mobile robot moves (hereinafter also referred to as the mobile environment), the mobile robot may acquire at each position of the movement route. By providing sensor motor data that can be used as teacher data and performing learning, the mobile robot autonomously follows the learning experience, that is, along the path through which sensor motor data as teacher data can be observed. Can move.

  In other words, a certain position on the learning path is given as the start position for starting movement, and the movement position is later on the learning path than the start position on the learning path as the goal position to end movement. Is given, the mobile robot can plan a route for moving from the start position to the goal position.

  However, in the mobile environment, the position on the path at the time of learning is not always given as the start position and the goal position, and the position after the start position on the path at the time of learning is given as the goal position. It is not necessarily done.

  That is, when the mobile robot moves autonomously, the current position becomes the start position, but the current position is not necessarily the position on the route at the time of learning.

  Furthermore, even if the start position and the goal position are positions on the route at the time of learning, a position that has passed before the start position that has been passed at the time of learning may be given as the goal position.

  Also, if the path from the start position to the goal position is redundant along the learning path, and if you make an unnecessarily detour to move from the start position to the goal position, do not make such a detour. It is desirable to have a route plan.

  As a conventional method of planning a route, for example, a movable area is obtained on a map of a moving environment, a graph is generated by using a line segment passing through the region as an arc, and a route is searched on the graph. There is a way to bring it back to the problem.

  As a method of searching for a route on the graph, a cost is set for each arc, and a route in which the sum of the costs of the arcs constituting the route is the smallest among the routes from the start position to the goal position is obtained. There is. As the arc cost, a distance on the map corresponding to the arc (distance between both ends of the arc) is used.

  However, in order to obtain the distance on the map corresponding to the arc, a map of the moving environment (and consequently the coordinates of the positions of both ends of the arc on the map) is required, and if no map is given It becomes difficult to find the distance on the map.

  Therefore, in preparation for the case where no map is given, it is desirable to adopt an index instead of the distance on the map corresponding to the arc as the cost of the arc.

  Therefore, in the learning device 10, connectivity representing the appropriateness of connection of the time series pattern learned by one other learning model is obtained after the time series pattern learned by one learning model.

  In the data generation device 20, the connectivity is adopted as the cost of the arc, and the route (shortest route) that minimizes the accumulated value of the cost (connection cost) by the graph route search algorithm such as the Viterbi algorithm or the Dijkstra method. A model sequence for generation as is searched.

  That is, in the data generation device 20, a value corresponding to connectivity is used as a connection cost, and an arrangement of learning models from the start point model to the end point model that minimizes the cumulative value of the connection cost is calculated as a generation model sequence.

  The connectivity used to calculate the model sequence for generation is one of the two learning models trained using the model learning data obtained by dividing the teacher data so that some of them overlap. Is the error between the last overlap part of the time series data generated by and the first overlap part of the time series data generated by the other learning model, and the model learning data used for learning one of the learning models And the model learning data used for learning of the other learning model do not depend on whether or not the training data is continuous in the teacher data (however, in the teacher data, the model learning data used for learning of the other learning model is used. If the data follows the model learning data used to learn one learning model, one learning model is With the Le, the connectivity model pair to the rear model the other learning models, becomes small).

  That is, in the teacher data, one learning model is generated even if the model learning data used for learning the other learning model is not the data following the model learning data used for learning one learning model. If the last overlap part of the time series data (time series pattern) and the first overlap part of the time series data generated by the other learning model are similar, one learning model is In addition, the connectivity of the model pair having the other learning model as the rear model is a small value indicating that it is appropriate to connect the rear model to the previous model.

  As a result, the arrangement of the learning models as the generation model sequence calculated based on the connectivity does not depend on the order of the model learning data used for learning the learning model on the teacher data.

  The learning model stores the time series pattern of the model learning data, which is a so-called fragment of the time series pattern of the teacher data, but the data generation device 20 uses the fragment and reuses the connection cost. It is possible to calculate a generating model sequence having a small accumulated value.

  That is, it is possible to calculate a generation model sequence that is not experienced during learning, for example, corresponding to a route that does not travel unnecessarily when moving from the start position to the goal position. In addition, for example, if a route moving in the opposite direction to the route experienced during learning is a route that reduces the cumulative value of connection costs, a generation model sequence corresponding to such a route can be calculated. it can.

  Furthermore, in the data processing device 20, since the arrangement of the learning models as the generation model sequence is calculated based on the connectivity, after the model generation data generated by a certain configuration model #k constituting the generation model sequence, It is ensured that it is appropriate to connect the model generation data generated by the immediately following configuration model # k + 1 (the waveform of the connected portion is similar).

  However, by calculating the model sequence for generation based on connectivity, when the model generation data generated by the configuration model # k + 1 is connected to the model generation data generated by the configuration model #k, There is no guarantee that the connection will be smooth.

  That is, when the arrangement of the configuration model of the generation model sequence calculated based on connectivity matches the order of the model learning data used for learning the configuration model on the teacher data, the configuration model When the model generation data generated by the constituent model # k + 1 is connected after the model generation data generated by #k, the connected portion becomes smooth.

  However, when the arrangement of the configuration model (learning model) of the generation model sequence calculated based on the connectivity does not match the order of the model learning data used for learning the configuration model on the teacher data When the model generation data generated by the configuration model # k + 1 is connected after the model generation data generated by the configuration model #k, the connection portion is not always smooth.

  Here, when the learning model stores the model learning data as a template as it is or when it is stored by function approximation without having an adjustable internal state, the time series data as it is stored (Model generation data) can only be generated.

  For this reason, when model generation data generated from a plurality of such learning models is connected, the connected portion is not always smooth.

  On the other hand, the data generation apparatus 20 employs an RNN having a context as an internal state that can store dynamics that evolve over time in the form of function approximation as a learning model. Further, in the data generation device 20, the last overlap portion of the model generation data generated by the configuration model #k and the first overlap portion of the model generation data generated by the configuration model # k + 1 connected later In order to reduce the error, an initial context of the RNN as a configuration model is determined, and the initial context (optimum initial context) is given to the configuration model to generate time series data.

  Therefore, when the model generation data generated by the configuration model # k + 1 is connected after the model generation data generated by the configuration model #k, the connected portion can be smoothed. Time series data can be generated.

[Generation time series data generated by the data generation device 20]
FIG. 19 shows time-series data as teacher data and generated time-series data generated using a learning model obtained by learning using the time-series data.

  FIG. 19A schematically shows a route (hereinafter also referred to as a teaching route) as teacher data.

The teaching path is one of the paths from positions P ₁ to P ₂ , and in FIG. 19A, model learning as _seven paths Q ₁ , Q ₂ , Q ₃ , Q ₄ , Q ₅ , Q ₆ , and Q ₇ is performed. Is divided into data. At the time of learning, the route Q _n is learned by the learning model #n.

  In FIG. 19, the overlap portion is not shown.

The learning model #n, which is an RNN, can be regarded as a function approximator that approximates the time evolution equation F (x, a) with the parameter a. Therefore, the learning model #n that has learned the route Q _n is hereinafter also expressed as F _n (x, a _n ).

  Here, the argument x of the time evolution equation F (x, a) represents input data, and the parameter a represents an initial value (initial context) of the internal state.

Further, in FIG. 19A, learning model F _n (x, a _n) parameters a _n are, for example, generates the learning model F _n (x, a _n) is the model generating data as possible matching path Q _n learned It represents the initial value of the internal state when it can be done.

Figure 19B by the data generating process of the data generating apparatus 20, learning model F ₁ (x, a) to F ₇ (x, a) the path of the generated time-series data generated using (hereinafter, both generation path Is schematically shown.

In FIG. 19B, generation path, albeit in a path from the position P ₁ to P _2, which is a different path from the taught path in FIG 19A.

That is, the generation path is configured by connecting model generation data as five paths Q ′ ₁ , Q ′ ₂ , Q ′ ₃ , Q ′ ₆ , and Q ′ ₇ in that order.

In FIG. 19B, the data generating apparatus 20, seven learning model F ₁ (x, a) learning model F ₄ from to no F ₇ (x, a), which generates a redundant path (x, a), and F ₅ Learning models F ₁ (x, a), F ₂ (x, a), F ₃ (x, a), F ₆ (x, a), F ₇ (x, a) excluding (x, a) A sequence is required as a model sequence for generation.

Further, in the generation of the generation path, the data generation apparatus 20 repeats the forward propagation and the reverse propagation of the overlap portion of the model generation data described in FIG. 12 and the like, thereby learning model F ₁ constituting the generation model sequence. _{(x, a), F 2} (x, a), F 3 (x, a), F 6 (x, a), F 7 (x, a) the overlapped portion of the model generation data generated from each of A parameter a for smooth connection is obtained.

In FIG. 19B, as parameters a overlapped portion is smoothly connected, the learning model F ₁ (x, a) the value a ₁ is the learning model F ₂ (x, a) is the value a _2, For the learning model F ₃ (x, a) the value a ' ₃ is for the learning model F ₆ (x, a), the value a' ₆ is for the learning model F ₇ (x, a), the value a ₇ is required for each.

The learning model F ₁ (x, a ₁₎ from the path Q _'1 are, learning model F ₂ (x, a ₂₎ from the path Q' is _2, for the learning model F ₃ (x, From a ′ ₃ ), the path Q ′ ₃ is from the learning model F ₆ (x, a ′ ₆ ), the path Q ′ ₆ is from the learning model F ₇ (x, a ₇ ), and the path Q ′ ₇ Are generated as model generation data.

In FIG. 19B, a path generated from the learning models F ₁ (x, a ₁ ), F ₂ (x, a ₂ ), and F ₇ (x, a ₇ ) in which the parameter a matches that in FIG. 19A. Q ′ ₁ , Q ′ ₂ , and Q ′ ₇ respectively correspond to the corresponding paths Q ₁ , Q ₂ , and Q ₇ in FIG. 19A.

On the other hand, in FIG. 19B, ₃ parameter a 'path Q is generated from ₍₃ different learning models F ₃ in the case of FIG 19A x, a)' is in Figure 19A, are different from the corresponding path Q _3.

That is, the route Q _{3 in} FIG. 19A has its start point side (side closer to the position P ₁ ) smoothly connected to the route Q ₂ , and the end point side (side closer to the position P ₂ ) It is adapted to smoothly connected to a path Q _4.

In contrast, the path Q in FIG. 19B _'3 the starting-point side, the same Q and path Q _2' is a point that is adapted to smoothly connect to _2, but consistent with the path Q _3, is the end point, It is different from the route Q _{3 in} that the route Q ′ ₆ is smoothly connected.

Further, in FIG. 19B, a path Q ′ ₆ generated from a learning model F ₆ (x, a ′ ₆ ) having a parameter a different from that in FIG. 19A is different from the corresponding path Q _{6 in} FIG. 19A.

That is, the route Q _{6 in} FIG. 19A has its start point side smoothly connected to the route Q ₅ and its end point side smoothly connected to the route Q ₇ .

In contrast, the path Q in FIG. 19B _'6 is the end point side, the same Q and path Q _7' is that adapted to smoothly connect to _7, but consistent with the path Q _6, is the starting point side, This is different from the route Q _{6 in} that the route Q ′ ₃ is smoothly connected.

  As described above, in the data generation device 20, redundant paths are excluded, and generation paths that are smoothly connected are generated.

[simulation result]
Next, a simulation performed by the inventor on the data processing apparatus shown in FIG. 1 will be described.

  In the simulation, the mobile robot performed navigation tasks.

  FIG. 20 shows an outline of a mobile environment in which a mobile robot performs a navigation task.

  As a moving environment, a two-dimensional plane in which light sources were installed and four sides were surrounded by walls was adopted. A mobile robot can move freely in a moving environment, but cannot move through a wall. In the mobile environment, there are walls that become obstacles in addition to the walls that surround the four sides.

  In addition, the mobile robot includes a distance sensor that senses the distance from the mobile robot to the wall (including both the wall surrounding the mobile environment and the wall as an obstacle in the mobile environment) in each of the eight surrounding directions. And the optical sensor which senses the intensity | strength of light, and the energy sensor which senses energy were mounted. Here, the energy is a physical quantity proportional to the maximum value of the light intensity in each of the eight directions output from the optical sensor.

The mobile robot includes a moving amount m _x in the horizontal direction (x-direction), the moving vector (m _x, m _y) is a vector representing the movement amount m _y in the vertical direction (y-direction), and as the motor data Given, the movement vector (m _x, m _y) moves only.

In the simulation, adopts above-described mobile robot, teacher data, as the present data, and the sensor motor data serving as the target data, the movement vector as a motor data (m _x, m _y), and, as a sensor data The distance sensor outputs the distances d ₁ , d ₂ , d ₃ , d ₄ , d ₅ , d ₆ , d ₇ , d _{8 for} each of the eight directions, and the light sensor outputs the light in each of the eight directions. intensity _{l 1, l 2, l 3} , l 4, l 5, l 6, l 7, l 8, and 19-dimensional vector having a component of the energy E of the energy sensor output (m _x, m _y, d ₁ , d ₂ , d ₃ , d ₄ , d ₅ , d ₆ , d ₇ , d ₈ , l ₁ , l ₂ , l ₃ , l ₄ , l ₅ , l ₆ , l ₇ , l ₈ , E) Adopted.

  The sensor motor data is observed from the mobile robot including a case where the human has manually moved the mobile robot.

  FIG. 21 is a plan view showing a moving environment adopted in the simulation.

  The moving environment is a two-dimensional plane surrounded by walls on all sides, and there are several light sources and several obstacle walls in the moving environment.

  In the learning process, the time series of sensor motor data of 15010 samples obtained by moving the mobile robot only 15010 times in the mobile environment was used as teacher data.

  The teacher data is divided into 40 samples of model learning data where the previous and next 10 samples overlap, and the resulting 500 model learning data are trained by the RNN as 500 learning models, respectively. It was.

  In FIG. 21, a three-digit number is a model ID for identifying 500 learning models. A line with a model ID (solid line, dotted line, thick line, thin line, etc.) represents the movement trajectory of the mobile robot when the model learning data learned by the learning model with that model ID is observed. ing.

  Here, the movement trajectory of the mobile robot when certain model learning data is observed is also referred to as a movement trajectory corresponding to the model learning data. Also, a learning model with a model ID of a three-digit number n is also described as a learning model #n.

  In FIG. 21, in order to make the model learning data learned by each learning model easy to understand, each time the learning model for learning the model learning data is switched, the movement locus corresponding to the model learning data is changed to the movement locus. The line types (solid line, dotted line, thick line, thin line, etc.) that represent are changed.

  Further, in FIG. 21, in order to avoid complication of the figure, the model learning data obtained by learning the first 80 learning models # 001 to # 080 out of 500 learning models # 001 to # 500 is shown. Only the corresponding movement trajectory is shown.

  FIG. 22 shows a mobile environment representing connectivity obtained by learning processing.

  In FIG. 22, a circle (circle) represents a certain position on the movement trajectory corresponding to the model learning data learned by the learning model, and corresponds to the learning model learned from the model learning data.

  In FIG. 22, only connectivity with high connectivity (connectivity with a value equal to or less than the threshold) is obtained by thresholding the connectivity obtained for 500 × 499 model pairs obtained from 500 learning models # 001 to # 500. Are extracted, and two learning models constituting a model pair having connectivity with high connectivity are connected by a dotted line.

  The mobile robot does not know the situation of the mobile environment, such as the arrangement of walls, but there is no value for connectivity that causes the mobile robot to move across the wall and cause the connection between learning models.

  In addition, the situation (structure) of the mobile environment is expressed (embedded) in a graph in which two learning models constituting a model pair having connectivity with high connectivity are connected by a dotted line.

  FIG. 23 shows a mobile environment representing a generation model sequence calculated using connectivity based on the Viterbi algorithm.

  In FIG. 23, as in FIG. 22, circles (◯) represent a certain position on the movement trajectory corresponding to the model learning data learned by the learning model, and the learning model obtained by learning the model learning data. Correspond.

  In FIG. 23, learning models (constitutive models) constituting the generation model sequence are connected by line segments.

  FIG. 24 shows a movement environment representing a movement locus corresponding to the model learning data learned by the learning model constituting the generation model sequence.

  In FIG. 24, a symbol string c: n consisting of a one-digit number c, a colon (:), and a three-digit number n indicates the order of learning models constituting the generation model sequence and the model ID of the learning model. Represents.

  That is, in the symbol string c: n, a one-digit number c indicates that the learning model is the c-th learning model from the beginning of the generation model sequence, and the three-digit number n following the colon after that is This represents that the c-th learning model from the top of the generation model sequence is learning model #n.

  Also, the line with the symbol string c: n represents a movement locus corresponding to the model learning data learned by the learning model #n.

  In FIG. 24, as in FIG. 21, in order to make the model learning data learned by each learning model easy to understand, each time the learning model for learning the model learning data is switched, the model learning data is handled. The movement trajectory is shown by changing the type of line representing the movement trajectory.

  In the simulation, the learning model # 049 is used as the start point model, the learning model # 277 is used as the end point model, the generation model sequence is calculated, and the learning models # 049, # 236, # 209, # 274, # 275, # A sequence of 276 and # 277 was obtained as a model sequence for generation.

  In order to reach learning model # 277, which is the end point model, from learning model # 049, which is the start point model, it is necessary to go through 227 learning models # 050 to # 276 according to experience with teacher data. is there.

  However, when the generation model sequence is calculated using connectivity, connectivity models with high connectivity and similar connectivity are connected to reduce the cumulative connection cost. As a result, it is possible to obtain a generation model sequence including a connection (a sequence of two learning models) that has no experience with teacher data.

  That is, the mobile robot has no experience with teacher data about the alignment of learning models # 049, # 236, the alignment of learning models # 236, # 209, and the alignment of learning models # 209, # 274, In FIG. 24, a generation model sequence including such a sequence is obtained.

  Here, learning models # 049, # 236 out of the arrangement of learning models # 049, # 236, the arrangement of learning models # 236, # 209, and the arrangement of learning models # 209, # 274 that have no experience with teacher data And the model pair as a sequence of learning models # 209 and # 274, the position on the teacher data of the model learning data learned by the previous model and the rear model that make up the model pair is the time series It is in order. On the other hand, for the model pair as an array of learning models # 236 and # 209, the position of the model learning data learned by the previous model and the rear model that make up the model pair on the teacher data is It is in the reverse order of the sequence order.

  As described above, in the calculation of the generation model sequence, since the learning model is connected so as to reduce the cumulative value of the connection cost, the position of the model learning data used for learning on the teacher data is A generation model sequence composed of a sequence of learning models that have no experience with teacher data, including a sequence of learning models that are in reverse order of time series, is obtained.

  As a result, the generation model sequence has a position (target data is observed) corresponding to the learning model # 277, which is the end point model, from the position corresponding to the learning model # 049, which is the starting point model (where the current data is observed). As a route that efficiently arrives at a position), a learning model is obtained in which a route shorter than the route where the teacher data is observed is obtained.

  FIG. 25 shows a movement environment representing a movement locus of the movement of the mobile robot so that generation time series data generated from the generation model sequence is observed.

  In FIG. 25, a symbol string c: n consisting of a single-digit number c, a colon (:), and a three-digit number n is similar to FIG. Indicates the model ID of the learning model.

  Further, in FIG. 25, a circle mark represents a start position where the mobile robot starts moving.

  As described in FIG. 24, the arrangement of the learning models # 049 and # 236, the arrangement of the learning models # 236 and # 209, and the arrangement of the learning models # 209 and # 274 in the generation model sequence depend on the teacher data. Not experienced, but smooth in switching from learning model # 049 to learning model # 236, switching from learning model # 236 to learning model # 209, and switching from learning model # 209 to learning model # 274 Therefore, it is possible to confirm that smooth generation time-series data is generated.

[Sharing model parameters]
In the above-described embodiment, a plurality of model learning data obtained by dividing the teacher data is learned using a plurality of learning models, and independent model parameters are obtained for each learning model. Model parameters can be shared for the model parameters.

  Hereinafter, sharing of model parameters will be described.

  FIG. 26 illustrates a configuration example of an embodiment of a learning apparatus that shares model parameters.

In FIG. 26, the learning apparatus includes a plurality of N learning modules 210 ₁ to 210 _N and a model parameter sharing unit 220.

Learning module _{210 i (i = 1,2, ···} , N) , the pattern input unit 211 _i, the model learning unit 212 _i, and is configured from the model storage unit 213 _i, using the model learning data, learning The model is learned, and model parameters of the learning model are obtained.

That is, the pattern input unit 211 _i, for model learning data is supplied for use in the learning of the stored learning model in the model storage unit 213 _i.

The pattern input unit 211 _i supplies the model learning data supplied thereto to the model learning unit 212 _i .

The model learning unit 212 _i uses the model learning data from the pattern input unit 211 _i to learn the learning model stored in the model storage unit 213 _i and obtains model parameters of the learning model.

The model storage unit 213 _i stores a learning model defined by model parameters for learning a pattern. That is, the model storage unit 213 _i stores model parameters of the learning model.

Here, the substance of the learning model stored in the model storage unit 213 _i is a model parameter of the learning model.

The model parameter sharing unit 220 performs a sharing process in which two or more learning modules out of the _N learning modules 210 ₁ to 210 _N share model parameters. When the model parameter sharing unit 220 performs the sharing process, two or more learning modules among the _N learning modules 210 ₁ to 210 _N share the model parameters.

Hereinafter, in order to simplify the description, the model parameter sharing unit 220 performs a sharing process in which all _N learning modules 210 ₁ to 210 _N share model parameters.

  Next, learning of a learning model performed by the learning device of FIG. 26 will be described with reference to the flowchart of FIG.

In step S211, the model learning unit 212 _i of the learning module 210 _i initializes the model parameters stored in the model storage unit 213 _i with, for example, random numbers, and the process proceeds to step S212.

In step S212, the learning module 210 _i waits for supply (input) of model learning data to be learned by the learning module 210 _i , and learning of the learning model using the model learning data, that is, The model parameters of the learning model are updated using the model parameters stored in the model storage unit 213 _i as initial values.

That is, in step S212, in the learning module 210 _i , the pattern input unit 211 _i supplies the model learning data supplied to the learning module 210 _i to the model learning unit 212 _i .

Further, in step S212, the model learning unit 212 _i updates the model parameters of the learning model stored in the model storage unit 213 _i using the model learning data from the pattern input unit 211 _i, and is obtained by the update. The stored contents of the model storage unit 213 _i are updated (overwritten) with the new model parameters thus obtained.

Here, the processes of steps S211 and S212 are performed by all of the _N learning modules 210 ₁ to 210 _N.

After step S212, the process proceeds to step S213, and the model parameter sharing unit 220 performs a sharing process in which all the _N learning modules 210 ₁ to 210 _N share the model parameter.

That is, the learning module 210 _i is (defined by a plurality of model parameters) having a plurality of model parameters. When attention is paid to, for example, the m-th model parameter among the plurality of model parameters of the learning module 210 _i , the model parameter sharing unit 220 determines the m-th model parameter of each of the _N learning modules 210 ₁ to 210 _N. based on, for correcting the m-th model parameter of the learning module 210 _1.

Further, the model parameter sharing unit 220 corrects the m-th model parameter of the learning module 210 ₂ based on the m-th model parameter of each of the _N learning modules 210 ₁ to 210 _N , and so on. The mth model parameter of each of the learning modules 210 ₃ to 210 _N is corrected.

As described above, the model parameter sharing unit 220 corrects the m-th model parameter of the learning module 210 _{i based on} the m-th model parameter of each of the N learning modules 210 ₁ to 210 _N , so that N each number of learning modules 210 ₁ to 210 _N m-th model parameters are to 210 ₁ to N learning modules all affected the m-th model parameter of 210 _N (to N learning module 210 no ₁ 210 _N each of the m-th model parameters, to 210 no ₁ N pieces of learning modules to affect all 210 _N m-th model parameters).

  In this way, all the model parameters of a plurality of learning modules are affected by each of the model parameters of the plurality of learning modules (each of the model parameters of the plurality of learning modules is It is the sharing of model parameters by multiple learning modules.

In step S213, the model parameter sharing unit 220 performs a sharing process on all of the plurality of model parameters stored in the model storage unit 213 _i of the learning module 210 _i , and uses the model parameters obtained by the sharing process. , model storage unit 213 ₁ to update the memory content of 213 _N.

  After step S213, the process proceeds to step S214, and the learning device in FIG. 26 determines whether the learning end condition is satisfied.

  Here, as the learning end condition in step S214, for example, it is possible to adopt that the number of times of learning, that is, the number of times that steps S212 and S213 are repeated has reached a predetermined number of times. . In addition, as the learning termination condition, when the true value of the output data to be output for the model learning data (output data for the input data when the model learning data is used as input data) is known In addition, it is possible to adopt that the error of the output data output from the learning model with respect to the model learning data is less than or equal to a predetermined value.

  If it is determined in step S214 that the learning termination condition is not satisfied, the process returns to step S212, and the same process is repeated thereafter.

  If it is determined in step S214 that the learning end condition is satisfied, the process ends.

  Next, FIG. 28 shows a configuration example of the learning apparatus in FIG. 26 when RNNPB (RNN with Parametric Bias) is adopted as a learning model.

Incidentally, in FIG. 28, is not shown in the learning module 210 _i of the pattern input section 211 _i and the model learning unit 212 _i.

The model storage unit 213 _i stores RNNPB (a plurality of model parameters that define RNNPB). Here, the RNNPB stored in the model storage unit 213 _i is hereinafter also referred to as RNNPB # i as appropriate.

  In FIG. 28, the RNNPB is composed of three layers: an input layer, a hidden layer (intermediate layer), and an output layer. Each of the input layer, the hidden layer, and the output layer is configured by an arbitrary number of units corresponding to neurons.

In RNNPB, the part of the unit is input units of the input layer, the model learning data x _t is input (supplied) as input data.

Further, the input layer, the PB unit is part of a unit other than the input unit model learning data x _t is input, PB (Parametric Bias) is input. According to PB, even if the same model learning data x _t is input to RNNPB in the same state, different output data x ^* _{t + 1} can be obtained by changing PB.

Output data output from some units in the output layer is fed back to the context unit, which is the rest of the units other than the input unit to which the model learning data x _t is input, as the context representing the internal state. Is done.

Here, when the model learning data xt at time _t is input to the input unit of the input layer, the PB and context at time t input to the PB unit and context unit of the input layer are respectively _expressed as PB _t and c _{Indicated as t} .

The hidden layer unit performs weighted addition using predetermined weights (weights) for model learning data x _t , PB _t , and context c _t input to the input layer, and the result of the weighted addition is an argument. And the result of the calculation is output to the output layer unit.

As described above, output data serving as the context c _{t + 1} at the next time t + 1 is output from some units in the output layer and fed back to the input layer. Further, from the remaining units of the output layer, for example, as output data to the model learning data x _t as input data, data for the model training x _t at the next time t + 1 of the model learning data x _{t +} predicted value x ^* _{t + 1} of ₁ is output.

  Here, in RNNPB, the input to the unit is weighted and added, and the weight (weight) used for this weighted addition is a model parameter of RNNPB. Weights as model parameters of RNNPB include, for example, weights from input units to hidden layer units, weights from PB units to hidden layer units, weights from context units to hidden layer units, and outputs from hidden layer units. There is a weight to the layer unit and a weight from the hidden layer unit to the context unit.

When RNNPB as described above is adopted as the learning model, the model parameter sharing unit 220 is provided with a weight matrix sharing unit 21 that allows the learning modules 210 ₁ to 210 _N to share weights as model parameters of the RNNPB.

  Here, there are a plurality of weights as model parameters of the RNNPB, but a matrix having the plurality of weights as components is called a weight matrix.

Weight matrix sharing unit 21 to RNNPB # 1 stored in the model storage unit 213 ₁ to 213 _N all weight matrix as a plurality of model parameters RNNPB # N, is shared by each of the learning modules 210 ₁ to 210 _N .

That is, if the weight matrix of RNNPB # i is expressed as w _i , the weight matrix sharing unit 21 sets the weight matrix w _i to the weight matrices w ₁ to w _N of the _N learning modules 210 ₁ to 210 _N, respectively. By performing correction based on all of them, a sharing process that affects all of the weight matrices w ₁ to w _N is performed on the weight matrix w _i .

Specifically, the weight matrix sharing unit 21 corrects the weight matrix w _i of RNNPB # i, for example, according to the following equation (3).

w _i = w _i + Δw _i
... (3)

Here, in equation (3), Δw _i is a correction component for correcting the weight matrix w _i and is obtained, for example, according to equation (4).

Δw _i = α _i Σβ _ij (w _j -w _i )
... (4)

  In Equation (4), Σ represents the total sum when the variable j is changed to an integer in the range of 1 to N.

The coefficient in the formula (4), beta _ij is representative of the weight matrix w _i of RNNPB # i, RNNPB # j ( j = 1,2, ···, N) the degree to which the influence of the weight matrix w _j of It is.

Therefore, the summation Σβ _ij (w _j -w _i ) on the right side of the equation (4) is the weight matrix w of RNNPB # 1 to RNNPB # N with respect to the weight matrix w _j of RNNPB # i using the coefficient β _ij as a weight. ₁ to w _N represents a weighted average value of deviations (differences), and α _i is a coefficient representing the degree of influence of the weighted average value Σβ _ij (w _j −w _i ) on the weight matrix w _i .

As the coefficients α _i and β _ij , for example, values larger than 0.0 and smaller than 1.0 can be adopted.

According to the equation (4), the smaller the coefficient α _i , the weaker the sharing (the influence of the weighted average value Σβ _ij (w _j −w _i ) on the weight matrix w _i becomes smaller), and the coefficient α as _i is large, so to speak, sharing becomes stronger.

Note that the method of correcting the weight matrix w _i is not limited to the equation (3), and can be performed, for example, according to the equation (5).

w _i = α ' _i w _i + (1-α' _i ) Σβ ' _ij w _j
... (5)

  In Equation (5), Σ represents the total sum when the variable j is changed to an integer in the range of 1 to N.

Further, in the equation (5), β _^'ij is the weight matrix w _i of RNNPB # i, represents the degree to which the influence RNNPB # j (j = 1,2, ···, N) of the weight matrix w _j of It is a coefficient.

Therefore, the summation Σβ ^′ _ij w _j in the second term on the right side of the equation (5) is the weighted average value of the weight matrices w ₁ to w _N of the RNNPB # 1 to RNNPB # N with the coefficient β ^′ _ij as the weight. Α ^′ _i is a coefficient representing the degree of influence of the weighted average value Σβ ^′ _ij w _j on the weight matrix w _i .

As the coefficients α ^′ _i and β ^′ _ij , for example, values larger than 0.0 and smaller than 1.0 can be adopted.

According to Equation (5), the larger the coefficient α ^′ _i , the weaker the sharing (the influence of the weighted average value Σβ ^′ _ij w _j received by the weight matrix w _i becomes smaller), and the smaller the coefficient α ^′ _i. The more it becomes, the stronger the sharing.

The plurality of learning modules 210 ₁ to 210 _N as a whole is excellent in scale extensibility that a new pattern can be easily learned.

Then, each of the plurality of learning modules 210 ₁ to 210 _N excellent in scale extensibility as a whole updates the model parameters of the plurality of learning modules 210 ₁ to 210 _N while sharing the model parameters (learning model). Learning), generalization characteristics obtained by learning performed by only one learning module can be obtained by the whole of the plurality of learning modules 210 ₁ to 210 _N , and as a result, there is a scalability as a whole. At the same time, a learning model having generalization characteristics can be obtained.

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

  Therefore, FIG. 29 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 on a hard disk 105 or a ROM 103 as a recording medium built in the computer.

  Alternatively, the program is stored temporarily on a removable recording medium 111 such as a flexible disk, a CD-ROM (Compact Disc Read Only Memory), a 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 111 can be provided as so-called package software.

  The program is installed in the computer from the removable recording medium 111 as described above, or transferred from the download site to the computer wirelessly via a digital satellite broadcasting artificial satellite, LAN (Local Area Network), The program can be transferred to a computer via a network such as the Internet, and the computer can receive the program transferred in this way by the communication unit 108 and install it in the built-in hard disk 105.

  The computer includes a CPU (Central Processing Unit) 102. An input / output interface 110 is connected to the CPU 102 via the bus 101, and the CPU 102 operates an input unit 107 including a keyboard, a mouse, a microphone, and the like by the user via the input / output interface 110. When a command is input by the equalization, a program stored in a ROM (Read Only Memory) 103 is executed accordingly. Alternatively, the CPU 102 also transfers from a program stored in the hard disk 105, a program transferred from a satellite or a network, received by the communication unit 108 and installed in the hard disk 105, or a removable recording medium 111 attached to the drive 109. The program read and installed in the hard disk 105 is loaded into a RAM (Random Access Memory) 104 and executed. Thus, the CPU 102 performs processing according to the above-described flowchart or processing performed by the configuration of the above-described block diagram. Then, the CPU 102 outputs the processing result from the output unit 106 configured with an LCD (Liquid Crystal Display), a speaker, or the like, for example, via the input / output interface 110, or from the communication unit 108 as necessary. Transmission and further recording on the hard disk 105 are performed.

  Here, in the present specification, the processing steps for describing a program for causing the computer to perform various processes do not necessarily have to be processed in time series in the order described in the flowcharts, but in parallel or individually. This includes processing to be executed (for example, parallel processing or processing by an object).

  Further, the program may be processed by one computer or may be distributedly processed by a plurality of computers. Furthermore, the program may be transferred to a remote computer and executed.

  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.

  That is, for example, in the calculation of connectivity and generation of generation time series data, as the overlap part of the model generation data, an L sample equal to the overlap part of the model learning data is adopted, and a plurality of less than the L sample is used. Samples can be employed.

It is a block diagram which shows the structural example of one Embodiment of the data processor to which this invention is applied. 3 is a block diagram illustrating a more detailed configuration example of a learning device 10. FIG. It is a figure which shows the structural example of RNN as a learning model. It is a figure explaining the division | segmentation of teacher data, and learning of the learning model using the data for model learning obtained by the division | segmentation. It is a figure explaining the method of calculating connectivity. It is a flowchart explaining a learning process. It is a flowchart explaining a connectivity calculation process. It is a flowchart explaining a connectivity calculation process. It is a flowchart explaining a connectivity calculation process. 3 is a block diagram illustrating a more detailed configuration example of a data generation device 20. FIG. It is a figure explaining the forward calculation performed for calculation of the model sequence for production | generation. It is a figure explaining the production | generation of the production | generation time series data using the model sequence for production | generation. It is a flowchart explaining a data generation process. It is a flowchart explaining the calculation process of the model sequence for production | generation. It is a flowchart explaining a time series data generation process. It is a flowchart explaining a time series data generation process. It is a flowchart explaining a time series data generation process. It is a flowchart explaining a time series data generation process. It is a figure which shows the time series data as teacher data, and the production | generation time series data produced | generated using the learning model which performed learning using the time series data. It is a figure which shows the outline | summary of the mobile environment where a mobile robot performs a navigation task. It is a top view which shows the movement environment employ | adopted by simulation. It is a figure which shows the movement environment showing connectivity obtained by the learning process. It is a figure which shows the movement environment showing the model sequence for production | generation calculated using connectivity. It is a figure which shows the movement environment showing the movement locus | trajectory corresponding to the data for model learning which the learning model which comprises the model sequence for production | generation learned. It is a figure which shows the movement environment showing the movement locus | trajectory which moved the mobile robot so that the production | generation time series data produced | generated from the model model for production | generation may be observed. It is a block diagram which shows the structural example of one Embodiment of the learning apparatus which shares a model parameter. It is a flowchart explaining the learning by a learning apparatus. It is a block diagram which shows the structural example of the learning apparatus at the time of employ | adopting RNNPB as a learning model. 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

DESCRIPTION OF SYMBOLS 10 Learning apparatus, 11 Teacher data preservation | save part, 12 Teacher data division | segmentation part, 13 Model learning data preservation | save part, 14 Learning part, 15 Model parameter preservation | save part, 16 Connectivity calculation part, 17 Connectivity preservation | save part, 20 Data generation apparatus, 21 Current data supply unit, 22 target data supply unit, 23 start point model selection unit, 24 end point model selection unit, 25 generation model sequence calculation unit, 26 time series data generation unit, 27 time series data output unit, 41 data allocation unit, 42 ₁ to 42 _N calculation unit, 51 model pair selection unit, 52 model parameter supply unit, 53, 54 recognition generation unit 7, 55 connectivity calculation unit, 61 current data distribution unit, 62 model parameter supply unit, 63 ₁ to 63 _N Recognition generation unit, 64 start point model determination unit, 71 target data Data distributing section, 72 model parameter supply section, 73 ₁ to 73 _N recognition generating unit, 74 an end point model determining unit 81 start the model ID supply unit, 82 an end point model ID supply unit, 83 sequence calculation unit, 91 sequence supplying section, 92 the model parameter supply section, 93 ₁ to 93 _N recognition generating unit 94 integrated generator, 101 bus, 102 CPU, 103 ROM, 104 RAM, 105 hard disk, 106 output unit, 107 input unit, 108 communication unit, 109 drive, 110 I / O interface, 111 removable recording medium, 210 ₁ to 210 _N learning module, 211 ₁ to 211 _N pattern input unit, 212 ₁ to 212 _N model learning unit, 213 ₁ to 213 _N model storage unit, 220 model parameter sharing unit, 221 Weight matrix sharing part

Claims (8)

The time series data is a learning model that divides time series data into a plurality of partially overlapping data and learns a time series pattern, and outputs it as model learning data used for learning a learning model having an internal state,
A plurality of model learning data obtained by dividing the time series data so as to assign one model learning data to one learning model, and assign the plurality of learning models to the learning model; One of the plurality of learning models after learning obtained by performing the learning of the time-series pattern by using the model learning data assigned to the learning model is time-series. Starting point model selecting means for selecting as a starting point model to be a starting point of a generating model sequence that is a sequence of the learning model used for generating data;
End point model selecting means for selecting one of the plurality of learning models as the end point model to be the end point of the generating model sequence;
For all of the plurality of learning models, a data string of the last part of the time series data generated by one learning model and a data string of the first part of the time series data generated by one other learning model A value representing an error is obtained by calculating the connectivity indicating the appropriateness of connection of the time series pattern learned by one other learning model after the time series pattern learned by one learning model. A value corresponding to the connectivity is set as a connection cost for connecting the other learning model after one learning model, and a cumulative value of the connection cost is minimized, from the start point model to the end point model. A generating model sequence calculating means for obtaining the sequence of the learning models as the generating model sequence;
For the learning model constituting the generation model sequence, the last partial data sequence of the time series data generated by the learning model and the first partial data of the time series data generated by the learning model connected later Time series data generating means for determining an initial value of the internal state of the learning model and giving the initial value to the learning model so as to reduce an error with a column, and generating time series data. Data processing device. The data processing apparatus according to claim 1, wherein the learning model is an RNN (Recurrent Neural Network). Data processing device
The time series data is a learning model that divides time series data into a plurality of partially overlapping data and learns a time series pattern, and outputs it as model learning data used for learning a learning model having an internal state,
A plurality of model learning data obtained by dividing the time series data so as to assign one model learning data to one learning model, and assign the plurality of learning models to the learning model; One of the plurality of learning models after learning obtained by performing the learning of the time-series pattern by using the model learning data assigned to the learning model is time-series. Select as the starting point model to be the starting point of the generating model sequence that is a sequence of the learning model used for data generation,
The other one of the learning models is selected as an end point model that is an end point of the generation model sequence,
For all of the plurality of learning models, a data string of the last part of the time series data generated by one learning model and a data string of the first part of the time series data generated by one other learning model A value representing an error is obtained by calculating the connectivity indicating the appropriateness of connection of the time series pattern learned by one other learning model after the time series pattern learned by one learning model. A value corresponding to the connectivity is set as a connection cost for connecting the other learning model after one learning model, and a cumulative value of the connection cost is minimized, from the start point model to the end point model. Obtaining the sequence of the learning models as the generating model sequence;
For the learning model constituting the generation model sequence, the last partial data sequence of the time series data generated by the learning model and the first partial data of the time series data generated by the learning model connected later A data processing method comprising the steps of: determining an initial value of the internal state of the learning model so as to reduce an error with a column; and providing the initial value to the learning model to generate time series data. The time series data is a learning model that divides time series data into a plurality of partially overlapping data and learns a time series pattern, and outputs it as model learning data used for learning a learning model having an internal state,
A plurality of model learning data obtained by dividing the time series data so as to assign one model learning data to one learning model, and assign the plurality of learning models to the learning model; One of the plurality of learning models after learning obtained by performing the learning of the time-series pattern by using the model learning data assigned to the learning model is time-series. Starting point model selecting means for selecting as a starting point model to be a starting point of a generating model sequence that is a sequence of the learning model used for generating data;
End point model selecting means for selecting one of the plurality of learning models as the end point model to be the end point of the generating model sequence;
For all of the plurality of learning models, a data string of the last part of the time series data generated by one learning model and a data string of the first part of the time series data generated by one other learning model A value representing an error is obtained by calculating the connectivity indicating the appropriateness of connection of the time series pattern learned by one other learning model after the time series pattern learned by one learning model. A value corresponding to the connectivity is set as a connection cost for connecting the other learning model after one learning model, and a cumulative value of the connection cost is minimized, from the start point model to the end point model. A generating model sequence calculating means for obtaining the sequence of the learning models as the generating model sequence;
For the learning model constituting the generation model sequence, the last partial data sequence of the time series data generated by the learning model and the first partial data of the time series data generated by the learning model connected later Time series data generating means for determining an initial value of the internal state of the learning model and giving the initial value to the learning model to generate time series data so as to reduce an error with a column. , A program to make a computer function. A learning model that divides time series data into a plurality of partially overlapping data and learns a time series pattern, and outputs a model learning data used for learning a learning model having an internal state; ,
A plurality of model learning data obtained by dividing the time series data so as to assign one model learning data to one learning model, and assign the plurality of learning models to the learning model; Learning means for learning the time-series pattern by using the model learning data assigned to the learning model;
For all of the plurality of learning models, a data string of the last part of the time series data generated by one learning model and a data string of the first part of the time series data generated by one other learning model Connectivity calculation means for calculating an error as a connectivity representing the appropriateness of connection of the time series pattern learned by another one of the learning models after the time series pattern learned by one of the learning models; Processing equipment. The data processing apparatus according to claim 5, wherein the learning model is an RNN (Recurrent Neural Network). Data processing device
The time series data is a learning model that divides time series data into a plurality of partially overlapping data and learns a time series pattern, and outputs it as model learning data used for learning a learning model having an internal state,
A plurality of model learning data obtained by dividing the time series data so as to assign one model learning data to one learning model, and assign the plurality of learning models to the learning model; The time-series pattern learning is performed using the model learning data assigned to the learning model,
For all of the plurality of learning models, a data string of the last part of the time series data generated by one learning model and a data string of the first part of the time series data generated by one other learning model A data processing method including a step of calculating an error as connectivity representing appropriateness of connection of the time series pattern learned by one other learning model after the time series pattern learned by one learning model. A learning model that divides time series data into a plurality of partially overlapping data and learns a time series pattern, and outputs a model learning data used for learning a learning model having an internal state; ,
A plurality of model learning data obtained by dividing the time series data so as to assign one model learning data to one learning model, and assign the plurality of learning models to the learning model; Learning means for learning the time-series pattern by using the model learning data assigned to the learning model;
For all of the plurality of learning models, a data string of the last part of the time series data generated by one learning model and a data string of the first part of the time series data generated by one other learning model Connectivity calculation means for calculating an error as connectivity representing the appropriateness of connection of the time series pattern learned by one other learning model after the time series pattern learned by one of the learning models, A program that allows a computer to function.

Images