JP2007305072A - Information processor, information processing method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To efficiently learn time-series data added, by making it difficult to change the weight coefficient of a learnt RNN. <P>SOLUTION: In the information processor 51, weight coefficient of RNN 71-n is learnt by giving a learning weight μn such that it makes difficult to change the weight coefficient to the RNN 71-n with a large use frequency FREQn in learning so far in additional learning processing of sensor motor signal as time-series data executed in RNN 71-1 to 71-n. The present invention is applicable to, for example, an information processor incorporated in a robot or the like. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an information processing device, an information processing method, and a program, and in particular, by making it difficult to change a weighting factor of a learned RNN, it is possible to efficiently learn added time-series data. The present invention relates to an information processing apparatus, an information processing method, and a program.

  The present applicant has previously proposed to generate time-series data according to the learning result using a recurrent neural network (see, for example, Patent Document 1).

  In this proposal, as shown in FIG. 1, the information processing apparatus basically includes a network in a lower hierarchy having recurrent neural networks (hereinafter referred to as RNN) 1-1 to 1-v, and RNN11- 1 to 11-v and an upper layer network.

  In the lower layer network, the outputs of RNNs 1-1 to 1-v are supplied to the synthesis circuit 3 via the corresponding gates 2-1 to 2-v, and synthesized.

  Also in the upper layer network, the outputs of the RNNs 11-1 to 11-v are supplied to the synthesis circuit 13 via the corresponding gates 12-1 to 12-v and synthesized. Then, on and off of the gates 2-1 to 2-v in the lower hierarchy are controlled based on the output of the synthesis circuit 13 in the upper hierarchy.

  In the information processing apparatus having the configuration shown in FIG. 1, time series data P1 to Pv are generated in RNNs 1-1 to 1-v in the lower hierarchy, respectively, and based on the output of the synthesis circuit 13 in the upper hierarchy, A time series in which a predetermined one of the RNNs 1-1 to 1-v is generated from the synthesis circuit 3 by turning on or off a predetermined one of the gates 2-1 to 2-v in the hierarchy of Any of the data P1 to Pv can be selectively output.

Thus, for example, as shown in FIG. 2, after the time series data P1 is generated for a predetermined time, the time series data P2 is generated for the next predetermined time, and further, the next predetermined time, The time series data can be generated by generating the time series data P1 again.
JP-A-11-126198

  However, when the time series data is learned by a plurality of RNNs as in the information processing apparatus of FIG. 1, after learning the predetermined time series data and adding new time series data to learn, If the new time-series data has similar characteristics to the trained time-series data, it is inherently different time-series data, but the weighted coefficient of the RNN that learned the learned time-series data with similar characteristics There was a problem that changed.

  The present invention has been made in view of such a situation, and makes it possible to efficiently learn added time-series data by making it difficult to change the weighting factor of the learned RNN. Is.

  An information processing apparatus according to an aspect of the present invention adds predetermined time series data to a plurality of recurrent neural networks after learning predetermined time series data for a plurality of recurrent neural networks. Learning weight determining means for determining a learning weight having a negative correlation with the frequency of use of each of the plurality of recurrent neural networks when the predetermined time-series data was previously learned. Each of the plurality of recurrent neural networks learns the new time-series data according to the learning weight determined by the learning weight determining means.

  According to an information processing method of one aspect of the present invention, after a plurality of recurrent neural networks learn predetermined time series data, new time series data is added to the plurality of recurrent neural networks. The learning weight having a negative correlation with the usage frequency of each of the plurality of recurrent neural networks when the predetermined time-series data was previously learned is determined, and the determined learning weight And learning the new time series data.

  The program according to one aspect of the present invention learns predetermined time-series data from a plurality of recurrent neural networks, and then adds new time-series data to the plurality of recurrent neural networks to learn. Determining a learning weight having a negative correlation with the usage frequency of each of the plurality of recurrent neural networks when the predetermined time series data was previously learned, and depending on the determined learning weight And causing the computer to execute a process including the step of learning the new time-series data.

  In one aspect of the present invention, when a plurality of recurrent neural networks are trained with predetermined time-series data, and then a plurality of recurrent-type neural networks are additionally trained with new time-series data. In addition, a learning weight having a negative correlation with the usage frequency of each of the plurality of recurrent neural networks when the predetermined time series data was previously learned is determined, and a new time is determined according to the determined learning weight. Series data is learned.

  According to one aspect of the present invention, it is possible to efficiently learn time-series data to be added by making it difficult to change a weighting factor of a learned RNN.

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

  An information processing apparatus according to an aspect of the present invention (for example, the information processing apparatus 51 in FIG. 3) causes a plurality of recurrent neural networks to learn predetermined time-series data, and then causes the plurality of recurrent neural networks to On the other hand, when new time series data is added and learned, the learning weight having a negative correlation with the frequency of use of each of the plurality of recurrent neural networks when the predetermined time series data was previously learned Learning weight determining means (for example, the control circuit 76 in FIG. 3), and each of the plurality of recurrent neural networks includes the new weight according to the learning weight determined by the learning weight determining means. Learn time-series data.

  An information processing method or program according to an aspect of the present invention (for example, the additional learning processing method of FIG. 11) causes a plurality of recurrent neural networks to learn predetermined time-series data, and then the plurality of recurrent neural networks. Learning to add negative time-series data to the network and learning negatively correlated with the frequency of use of each of the plurality of recurrent neural networks when the predetermined time-series data was previously learned Is determined (for example, step S102 in FIG. 11), and the new time series data is learned (for example, step S103 in FIG. 11) according to the determined learning weight.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 3 shows a configuration example of an information processing apparatus to which the present invention is applied.

  The information processing apparatus 51 in FIG. 3 is incorporated in, for example, a robot. The robot in which the information processing device 51 is incorporated includes at least a sensor that detects an object to be visually recognized and a motor (none of which is shown) that is driven to move the robot. A sensor motor signal that is a signal from the motor is supplied to the information processing apparatus 51.

  The information processing apparatus 51 includes a lower time series prediction generator 61, an upper time series prediction generator 62, and a gate signal converter 63, and learning processing for learning time series data given as teacher data, and learning Depending on the result, a generation process for generating (reproducing) time-series data for the input is executed.

  In the present embodiment, an example will be described in which the information processing apparatus 51 learns and generates a behavior sequence that is a series of operations performed by a humanoid robot.

  In the following example, the information processing apparatus 51 learns three action sequences A, B, and C.

  The action of humanoid robot as action sequence A is as follows.A robot with both arms extended to the left and right as an initial state visually recognizes a square object placed on the table in front of the user, grabs the object with both hands, and This is an operation of lifting the height and placing it again on the table a plurality of times, and then returning both arms to the initial position (hereinafter also referred to as the home position).

  From the initial state, the humanoid robot moves as an action sequence B by visually recognizing a square object placed on the table in front of you, touching the object with the right hand, returning to the home position, touching the object with the left hand, and moving to the home position. The operation of returning, that is, the operation of alternately touching the object with one hand is performed a plurality of times.

  The action of the humanoid robot as the action sequence C is an action of visually recognizing a square object placed on a table in front of the eyes from the initial state, touching the object once with both hands, and returning to the home position.

  The information processing apparatus 51 learns and generates signals from a sensor (for example, a visual sensor) and a motor when executing each of the action sequences A to C as described above.

  The low-order time series prediction generator 61 includes N recurrent neural networks (hereinafter referred to as RNNs) 71-1 to 71-N, and a gate 72 disposed in a subsequent stage of the RNNs 71-1 to 71-N. -1 to 72-N, a synthesis circuit 73, an arithmetic circuit 74, a memory 75, and a control circuit 76. Note that RNNs 71-1 to 71-N are simply referred to as RNN 71 when it is not necessary to distinguish between them. The same applies to the other gates 72 and the like.

  The lower time series prediction generator 61 receives sensor motor signals from the center and motor provided in the humanoid robot. Here, the sensor motor signal input to the low-order time series prediction generator 61 at time t is represented as sm (t).

  The low-order time series prediction generator 61 predicts the sensor motor signal sm (t + 1) at time t + 1 according to the previously learned result with respect to the sensor motor signal sm (t) at time t input thereto. ,Output.

  Specifically, the RNN 71-n (n = 1, 2,..., N) receives the sensor motor signal sm (t) at the input time t at the time t + 1 according to the result learned in advance. A sensor motor signal sm (t + 1) is generated and output to the gate 72-n.

  By the way, the action sequence can be considered to be composed of a collection (continuous) of various action parts (motion primitives). For example, the above-described behavior sequence A is considered to be a collection of behavioral parts such as visually recognizing an object, bringing both arms close to the object (until grasping), lifting the object, lowering the lifted object, and returning to the home position. be able to. Each of the RNNs 71-1 to 71-N exclusively learns time series data of sensor motor signals corresponding to one action component.

Therefore, since the behavioral components learned by the RNNs 71-1 to 71-N are different, the same sensor motor signal sm (t) is input to each of the RNNs 71-1 to 71-N. The sensor motor signals sm (t + 1) output from −1 to 71-N are different. Here, the sensor motor signal sm (t + 1) output by the RNN 71-n is represented as a sensor motor signal sm _n (t + 1).

In addition to the sensor motor signal sm _n (t + 1) at the time t + 1 from the RNN 71-n, the gate 72-n disposed at the subsequent stage of the RNN 71-n receives the gates 72-1 to 72- from the gate signal conversion unit 63. A gate signal gate [N] = {g ₁ , g ₂ ,..., G _N }, which is a control signal for N open / close states, is supplied. As will be described later, the total sum of the gate signals g _n constituting the gate signal gate [N] is 1 (Σg _n = 1).

Gate 72-n in response to the gate signal g _n, to open and close the output of the sensor motor signal sm _n (t + 1) from RNN71-n. That is, the gate 72-n outputs g _n × sm _n (t + 1) to the synthesis circuit 73 at time t + 1.

  The synthesis circuit 73 synthesizes the outputs from the gates 72-1 to 72-N, and outputs the result of the synthesis as a sensor motor signal sm (t + 1) at time t + 1. That is, the synthesis circuit 73 outputs a sensor motor signal sm (t + 1) expressed by the following equation (1).

When learning the time series data of the sensor motor signal, the arithmetic circuit 74 outputs the sensor motor signal sm ₁ (t + 1) at time t + 1 output from each of the RNNs 71-1 to 71-N with respect to the sensor motor signal sm (t) at time t. ) To sm _N (t + 1) and the prediction error errorL ^{t + 1} [N] = {errorL ^{t +} between the teacher sensor motor signal sm ^* (t + 1) at time t + 1 given to the lower time series prediction generator 61 as teacher data ¹ ₁ , errorL ^{t + 1} ₂ ,..., ErrorL ^{t + 1} _N } are calculated. Note that the prediction error errorL ^{t + 1} [N] is calculated not only as an error at time t + 1 but also as an error taking into account the past L steps from time t + 1, as expressed by equation (16) described later.

The prediction error errorL ^{t + 1} _n of RNN 71-n at time t + 1 calculated by the arithmetic circuit 74 is supplied to the memory 75 and stored therein.

In the arithmetic circuit 74, the calculation of the prediction error errorL ^{t + 1} [N] is repeated in time series and stored in the memory 75, whereby the memory 75 stores the prediction error time series data errorL [N] for the teacher data. Is memorized. The time series data errorL [N] of the prediction error is supplied to the upper time series prediction generator 62. The arithmetic circuit 74 normalizes the prediction error time series data errorL [N] with respect to the teacher data to a value in the range of 0 to 1, and then outputs the normalized data.

As described above, the memory 75 stores the time series data errorL [N] of the prediction error for the teacher data. The memory 75 also stores the usage frequencies FREQ _{1 to} FREQ _N of the RNNs 71-1 to 71-N. The usage frequencies FREQ _{1 to} FREQ _N of the RNNs 71-1 to 71-N will be described later with reference to FIG.

  The control circuit 76 controls each unit of the lower time series prediction generator 61 such as the RNNs 71-1 to 71-N, the arithmetic circuit 74, and the memory 75.

  On the other hand, the high-order time-series prediction generator 62 includes a single continuous-time RNN (Continuous Time RNN: hereinafter referred to as CTRNN) 81.

  The CTRNN 81 of the upper time series prediction generator 62 estimates (predicts) how much prediction error the RNNs 71-1 to 71-N of the lower time series generator 61 generate at the time of generation, and outputs them.

That is, the CTRNN 81 learns by using the time series data errorL [N] of the prediction errors of the RNNs 71-1 to 71-N as teacher data, and based on the learned result, the estimated prediction of the RNNs 71-1 to 71-N The error errorPredH [N] = {errorPredH ₁ , errorPredH ₂ ,..., ErrorPredH _N } is generated and output. Here, the estimated prediction error errorPredH [N] at time t is assumed to be errorPredH ^t [N] = {errorPredH ^t ₁ , errorPredH ^t ₂ ,..., ErrorPredH ^t _N }.

  Also, CTRNN 81 is given a task ID as a task switching signal for switching which estimated prediction error errorPredH [N] of behavior sequences A to B is to be output.

The gate signal converter unit 63, using the soft max (softmax) function, the estimated prediction at time t error errorPredH ^t [N], the gate signal ^{gate t [N] = {g} t 1, g t 2, ··· , G ^t _N }, and outputs the converted result to the gates 72-1 to 72-N.

The gate signal g ^t _n to the gate 72-n at time t is expressed by the following formula (2).

  According to Equation (2), nonlinear conversion is performed such that a small prediction error has a large value and a large prediction error has a small value. As a result, control is performed in the gates 72-1 to 72-N of the lower time series generator 61 such that the smaller the prediction error, the larger the gate opens, and the larger the prediction error, the smaller the gate opens. become.

In the information processing apparatus 51 configured as described above, the higher-order time-series prediction generator 62 is an estimated prediction that is an estimated value of a prediction error generated by the RNNs 71-1 to 71-N of the lower-order time series generator 61 at the time of generation. An error errorPredH [N] is output, and the estimated prediction error errorPredH [N] is converted into a gate signal gate [N] that controls the open / closed state of the gates 72-1 to 72-N. Then, the output signals sm ₁ (t + 1) to sm _N (output from the RNNs 71-1 to 71-N output from the gates 72-1 to 72-N whose open / close states are controlled, expressed by the above-described equation (1). The sum of (t + 1) is supplied to the sensor and motor of the humanoid robot as a sensor motor signal sm (t + 1) at time t + 1.

  The upper time series prediction generator 62 converts the estimated prediction error errorPredH [N], which is the output of the upper time series prediction generator 62, into the gate signal gate [N] in the subsequent gate signal conversion unit 63. It can also be said that the gates 72-1 to 72-N to be opened (largely) at the time t are predicted.

  FIG. 4 shows a detailed configuration example of the RNN 71-n.

  As shown in FIG. 4, the RNN 71-n includes an input layer 101, an intermediate layer (hidden layer) 102, and an output layer 103. The input layer 101 has a predetermined number of nodes 111, and the intermediate layer The (hidden layer) 102 has a predetermined number of nodes 112, and the output layer 103 has a predetermined number of nodes 113.

  To the node 111 of the input layer 101, the sensor motor signal sm (t) at time t and the node 113 which is a part of the output layer 103 are output from the node 113 of the output layer 103 at time t-1 immediately before time t. Data fed back as the context c (t) representing the internal state is input.

  The node 112 of the intermediate layer 102 performs a weighted addition process that multiplies the data input from the node 111 of the input layer 101 and the weighting coefficient between the node 111 obtained in advance by learning, and outputs the calculation result. Output to the node 113 of the output layer 103.

The node 113 constituting the output layer 103 performs an operation of weighted addition processing for multiplying and summing the data input from the node 112 of the intermediate layer 102 and the weighting coefficient between the node 112 obtained in advance by learning. Some nodes 113 constituting the output layer 103 output the calculation result as a sensor motor signal sm _n (t + 1) at time t + 1. The other part of the nodes 113 constituting the output layer 103 feeds back the calculation result to the node 111 of the input layer 101 as the context c (t + 1) at time t + 1.

As described above, the RNN 71-n detects the sensor at the time t + 1 with respect to the input sensor motor signal sm (t) at the time t by the weighted addition process using the weighting coefficient between the nodes obtained by learning in advance. The motor signal sm _n (t + 1) is predicted and output.

  Note that a BPTT (Back Propagation Through Time) method is employed in learning for obtaining a weighting coefficient between nodes. The BPTT method is an RNN learning algorithm with a context loop, and it applies the back-propagation (BP) method in a normal hierarchical neural network by spatially expanding the state of signal propagation over time. is there. The same applies to the case where the weighting coefficient is obtained in CTRNN 81 described later.

  FIG. 5 shows a detailed configuration example of CTRNN adopted as CTRNN81.

  The CTRNN 141 in FIG. 5 includes an input layer 151, an intermediate layer (hidden layer) 152, an output layer 153, and arithmetic units 154 and 155.

  The input layer 151 includes input nodes 160-i (i = 1,..., I), parameter nodes 161-r (r = 1,..., R), and context input nodes 162-k (k = 1). ,..., K), and the intermediate layer 152 has hidden nodes 163-j (j = 1,..., J). The output layer 153 includes output nodes 164-i (i = 1,..., I) and context output nodes 165-k (k = 1,..., K).

  Note that it is not necessary to distinguish between the input node 160-i, the parameter node 161-r, the context input node 162-k, the hidden node 163-j, the output node 164-i, and the context output node 165-k. Are simply referred to as an input node 160, a parameter node 161, a context input node 162, a hidden node 163, an output node 164, and a context output node 165.

The CTRNN 141 learns to predict and output the state vector x ^u (t + 1) at time t + 1 with respect to the state vector x ^u (t) input at time t. The CTRNN 141 has a regression loop called a context loop representing the internal state of the network, and can learn the time evolution law of the target time-series data by performing processing based on the internal state.

The state vector x ^u (t) at time t supplied to the CTRNN 141 is input to the input node 160. A parameter tsdata ^u is input to the parameter node 161. The parameter tsdata ^u is data that identifies the type (time-series data pattern) of the state vector x ^u (t) supplied to the CTRNN 141, and the CTRNN 81 is data that identifies the action sequence. Although the parameter tsdata ^u is a fixed value, it can be considered that the same value is continuously input. Therefore, the data (vector) input to the parameter node 161 at time t is defined as the parameter tsdata ^u (t). To do.

The input node 160-i receives data x ^u _i (t) that is the i-th element constituting the state vector x ^u (t) at time t. The parameter node 161-r receives data tsdata ^u _r (t) which is the r-th element constituting the parameter tsdata ^u (t) at time t. Furthermore, data c ^u _k (t), which is the k-th element constituting the internal state vector c ^u (t) of CTRNN 141 at time t, is input to the context input node 162-k.

When data x ^u _i (t), tsdata ^u _r (t), and c ^u _k (t) are input to the input node 160-i, the parameter node 161-r, and the context input node 162-k, respectively. , Input node 160-i, parameter node 161-r, and context input node 162-k output data x _i (t), tsdata _r (t), and c _k (t), respectively, 3), represented by formula (4), and formula (5).

The function f in the equations (3) to (5) is a differentiable continuous function such as a sigmoid function, and the equations (3) to (5) include the input node 160-i, the parameter node 161-r, and Data x ^u _i (t), tsdata ^u _r (t), and data c ^u _k (t) input to each of the context input nodes 162-k are activated by the function f, and the data x _i (t) , Tsdata _r (t), and data c _k (t) are output from the input node 160-i, the parameter node 161-r, and the context input node 162-k. The superscript ^{u of} the data x ^u _i (t), tsdata ^u _r (t), and c ^u _k (t) represents the internal state of the node before being activated (the same applies to other nodes). ).

The data h ^u _j (t) input to the hidden node 163-j is a weight coefficient w ^h _ij representing the weight of the connection between the input node 160-i and the hidden node 163-j, the parameter node 161-r and the hidden node 163. Using the weight coefficient w ^h _jr representing the weight of the connection of −j and the weight coefficient w ^h _jk representing the weight of the connection between the context input node 162 -k and the hidden node 163 -j, expressed by Expression (6) The data h _j (t) output from the hidden node 163-j can be expressed by Expression (7).

  Note that Σ in the first term on the right side of Equation (6) indicates that addition is performed for all of i = 1 to I, and Σ in the second term indicates that addition is performed for all of r = 1 to R. Σ in the third term represents addition for all of k = 1 to K.

Similarly, the data y ^u _i which is input to the output node 164-i (t), the output node 164-i outputs data y _i (t), and the data to be input to the context output nodes 165-k o ^u _k (t) and data o _k (t) output from the context output node 165-k can be expressed by the following equations.

In formula (8), w ^y _ij is a weighting coefficient that represents the weight of the connection between the hidden node 163-j and the output node 164-i, and Σ represents that all j = 1 to J are added. Also, w ^o _jk in equation (10) is a weighting coefficient that represents the weight of the connection between the hidden node 163-j and the context output node 165-k, and Σ is added for all of j = 1 to J. To express.

Calculation unit 154, the data y _i (t) to the output node 164-i outputs the difference △ x ^u _i between the time data t x ^u _i (t) at time t + 1 of the data _{^{x u i (t + 1)}} ( t + 1) is obtained from equation (12), and data x ^u _i (t + 1) at time t + 1 is calculated and output from equation (13).

  Here, α and τ represent arbitrary coefficients.

Therefore, when data x ^u _i (t) at time t is input to CTRNN 141, data x ^u _i (t + 1) at time t + 1 is output from the calculation unit 154 of CTRNN 141. Further, the data x ^u _i (t + 1) at time t + 1 output from the calculation unit 154 is also supplied (feedback) to the input node 160-i.

The computing unit 155 determines the difference Δc ^u between the data c ^u _k (t) at time t and the data c ^u _k (t + 1) at time t + 1 from the data o _k (t) output from the context output node 165 -k. _k (t + 1) is obtained by Expression (14), and data c ^u _k (t + 1) at time t + 1 is calculated and output by Expression (15).

The data c ^u _k (t + 1) at time t + 1 output from the calculation unit 155 is fed back to the context input node 162-k.

Expression (15) adds the data o _k (t), which is the output of the context output node 165-k, to the internal state vector c ^u (t) representing the current internal state of the network, weighted by the coefficient α (predetermined). This means that the internal state vector c ^u (t + 1) of the network at the next time t + 1 is obtained. In this sense, CTRNN 141 in FIG. 5 is a continuous-time RNN. I can say that.

As described above, when data x ^u (t) and data c ^u (t) at time t are input, CTRNN 141 generates data x ^u (t + 1) and data c ^u (t + 1) at time t + 1. since the process of outputting sequentially the weighting factor _{^{w h ij, w h ir,}} w h jk, When w ^y _ij, and w ^o _jk is obtained by learning, the input data to be input to the input node 160 x ^u (t) initial value x ^u (t ₀ ) = X 0, parameter tsdata ^u input to parameter node 161, initial value c ^u (t ₀ ) of context input data c ^u (t) input to context input node 162 ) = C0, time series data can be generated.

The CRTNN 141 shown in FIG. 5 is adopted as the CRTNN 81 in FIG. 3, errorL [N] is given to the input node 160 of the CRTNN 141, and the task ID is given to the parameter node 161. Accordingly, the number I of input nodes 160 in FIG. 5 matches the number N of RNNs 71 in the lower time series prediction generator 61. For example, a random predetermined value is given as the initial value c ^u (t ₀ ) = C _{0 of} the context input data c ^u (t) input to the context input node 162.

  Next, with reference to the flowchart of FIG. 6, the learning process of the time series data of the sensor motor signal corresponding to the action sequence in the lower time series prediction generator 61 will be described.

  First, in step S1, the control circuit 76 of the lower time series prediction generator 61 reads input data at a predetermined time supplied as teacher data. The input data here is a sensor motor signal as described above. For example, it is assumed that the sensor motor signal sm (t) at time t is read. The read sensor motor signal sm (t) at time t is supplied by the control circuit 76 to each of the N RNNs 71-1 to 71-N constituting the lower time series prediction generator 61.

In step S2, the RNN 71-n (n = 1, 2,..., N) of the lower time series prediction generator 61 detects the sensor motor signal at time t + 1 with respect to the sensor motor signal sm (t) at time t. sm _n (t + 1) is calculated.

In step S2, the arithmetic circuit 74 calculates a prediction error errorL ^{t + 1} _n of the RNN 71-n. Specifically, the arithmetic circuit 74 calculates a prediction error with respect to the sensor motor signal for the past L time steps from time t + 1, expressed by the equation (16), as a prediction error errorL ^{t + 1} _n .

In equation (16), sm _{n, i ′} (T) is the number of nodes 113 (FIG. 4) in the output layer 103 of the RNN 71-n that outputs the sensor motor signal sm (T) at time T. The sensor motor signal output from the i'th node 113 is represented, and sm ^* _{n, i '} (T) represents the sensor motor signal as teacher data for the sensor motor signal.

According to equation (16), the sensor motor signal sm _{n, i ′} (T) and the teacher data sm ^* of the i′-th node 113 of the output layer 103 of the RNN 71-n from time T = t + 1−L to t + 1 ^. The sum of errors from _{n, i ′} (T) is the prediction error errorL ^{t + 1} _n of RNN 71-n at time t + 1. When the past sensor motor signal does not include L time steps, the prediction error errorL ^{t + 1} _n is obtained using only data for existing time steps.

In step S <b ^{> 3} , the arithmetic circuit 74 supplies the prediction error errorL ^{t + 1} _n of the RNN 71 -n at time t + 1 to the memory 75. Accordingly, n prediction errors errorL ^{t + 1} _{1 to} errorL ^{t + 1} _{N of} RNNs 71-1 to 71- _N are supplied to the memory 75, and the memory 75 stores the prediction errors errorL ^{t + 1} [N] = Store {errorL ^{t + 1} ₁ , errorL ^{t + 1} ₂ ,..., ErrorL ^{t + 1} _N }. If it is determined that the process in step S7, which will be described later, is NO, the process in step S3 is repeated for a predetermined time step, so that the memory 75 stores time series data errorL [N] of the prediction error for the teacher data. .

In step S4, the control circuit 76 calculates the learning weight γ _n of the RNN 71-n corresponding to the prediction error errorL ^{t + 1} _n . Specifically, the control circuit 76 calculates the learning weight γ _{n according} to Expression (17) using a softmax function (softmax function).

In step S5, the control circuit 76 updates the weight coefficient w _{ab, n} of the RNN 71-n by a BPTT (Back Propagation Through Time) method. Here, the weighting factor w _{ab, n} is the weighting factor between the node 111 of the input layer 101 of the RNN 71-n and the node 112 of the intermediate layer 102, or between the node 112 of the intermediate layer 102 of the RNN 71-n and the output layer 102. The weight coefficient with the node 113 is represented.

RNN71-n weighting coefficient w _{ab, and} in updating _n, the weighting coefficient w _ab of RNN71-n in accordance with the learning weights gamma _n calculated in step _{S4, n} is calculated. Specifically, the s + 1-th weight coefficient w _{ab, n} (s + 1) is obtained from the s- _th weight coefficient w _{ab, n} (s) in the repetitive calculation of the BPTT method by the following equations (18) and (19). be able to.

In Expression (18), η ₁ represents a learning coefficient, and α ₁ represents an inertia coefficient. In Equation (18), _{Δwab, n} (s) when s = 1 is set to 0.

In step S <b> 6, the control circuit 76 supplies the usage frequencies FREQ _{1 to} FREQ _N of the RNNs 71-1 to 71- _N to the memory 75. The memory 75 stores the usage frequencies FREQ _{1 to} FREQ _N of the supplied RNNs 71-1 to 71-N. As the learning weight γ _n is larger in step S5 described above, the weight coefficient w _{ab, n} of the RNN 71-n is updated, and the RNN 71-n is used. Accordingly, the control circuit 76, for example, learning weights gamma _n is to count up the use frequency FREQ _n of RNN71-n is greater than or equal to a predetermined value. The usage frequencies FREQ _{1 to} FREQ _N are used in additional learning described later with reference to FIG.

  In step S7, the control circuit 76 of the lower time series prediction generator 61 determines whether or not the supply of input data has been completed.

  If it is determined in step S7 that the supply of input data has not ended, that is, if input data at the next time of the input data supplied in step S1 is supplied, the process returns to step S1, and thereafter The process is repeated.

  On the other hand, if it is determined in step S7 that the supply of input data has ended, the learning process ends.

  Next, learning of time series data of prediction errors in the CRTNN 81 of the upper time series prediction generator 62 will be described.

When a humanoid robot equipped with the information processing device 51 learns a plurality of action sequences, weight coefficients w ^h _ij , w ^h _jr , and w between nodes of the input layer 151 and the intermediate layer 152 obtained as a result of learning ^h _jk and the weight coefficients w ^y _ij and w ^o _jk between the nodes of the intermediate layer 152 and the output layer 153 need to be values that can correspond to all action sequences.

Therefore, in the learning process, learning of time series data corresponding to a plurality of action sequences is performed simultaneously. That is, in the learning process, the same number of CTRNNs 141 (FIG. 5) as the number of action sequences to be learned are prepared, and the weight coefficients w ^h _ij , w ^h _jr , w ^h _jk , w ^y _ij , and w ^o _jk for each action sequence. CTRNN81 used in the generation process is obtained by repeatedly executing the process of obtaining each of the average values and setting the average value thereof as one weighting coefficient w ^h _ij , w ^h _jr , w ^h _jk , w ^y _ij , and w ^o _jk. weight coefficients w ^h _{ij ^{of, w h jr, w h jk}} , w y ij, and w ^o _jk is required.

  FIG. 7 is a flowchart of the learning process of the high-order time-series prediction generator 62 that learns time-series data of Q prediction errors corresponding to Q action sequences. In the present embodiment, since there are three action sequences to be learned, action sequences A, B, and C, Q = 3.

  First, in step S <b> 31, the upper time series prediction generator 62 reads time series data errorL [N] of Q prediction errors as teacher data from the memory 75 of the lower time series prediction generator 61. Then, the upper time series prediction generator 62 supplies the read Q pieces of time series data errorL [N] to the Q pieces of CRTNN 141, respectively.

  In step S32, the upper time-series prediction generator 62 reads a task ID for identifying each of the Q action sequences. In this embodiment, a task ID for identifying each of the three action sequences A, B, and C is read. Then, the upper time series prediction generator 62 supplies the task ID for identifying the action sequence A to the CRT 141 that has supplied the teacher data for the action sequence A, and the action RT to the CRT 141 that has supplied the teacher data for the action sequence B. A task ID for identifying the sequence B is supplied, and a task ID for identifying the behavior sequence C is supplied to the CRT 141 that has supplied the teacher data of the behavior sequence C.

  In step S33, the higher-order time-series prediction generator 62 substitutes 1 for a variable s representing the number of learning times.

In step S34, the higher-order time-series prediction generator 62 uses the BPTT method in the CTRNN 141 corresponding to each of the Q pieces of time-series data to use the weighting coefficient w ^h _ij (between the nodes of the input layer 151 and the intermediate layer 152. s), w ^h _jr (s), and w ^h _jk (s) error amounts δw ^h _ij , δw ^h _jr , and δw ^h _jk, and a weighting factor w ^y between each node of the intermediate layer 152 and the output layer 153. The error amounts δw ^y _ij and δw ^o _jk of _ij (s) and w ^o _jk (s) are calculated. Here, the error amounts δw ^h _ij , δw ^h _jr , δw ^h _jk , δw ^y obtained by using the BPTT method in the CTRNN 141 to which the q (= 1,..., Q) -th time series data is input. _ij and δw ^o _jk are expressed as error amounts δw ^h _{ij, q} , δw ^h _{jr, q} , δw ^h _{jk, q} , δw ^y _{ij, q} , and δw ^o _{jk, q} , respectively.

In the calculation using the BPTT method in step S34, the upper time series prediction generator 62 calculates the error amount δc ^u _k (t + 1) of the data c ^u _k (t + 1) of the context input node 162-k at time t + 1, when backpropagated error amount .delta.o _k (t) of time t context output nodes 165-k of the data o _k (t), by dividing any positive coefficients m, to adjust the time constant of the context data .

That is, the upper time series prediction generator 62, the error amount .delta.o _k (t) of the data o _k context output nodes 165-k at time t (t), the time t + 1 of the context input node 162-k of the data c ^u _{This is obtained} by the equation (20) using the error amount δc ^u _k (t + 1) of _k (t + 1).

  By adopting the equation (20) in the BPTT method, it is possible to adjust the influence level of the context data representing the internal state of the CTRNN 141 one step ahead.

In step S35, the upper time series prediction generator 62, each of the input layer 151 and the weighting coefficients w ^h _ij between nodes of the intermediate layer 152, w ^h _jr, and w ^h _jk and an intermediate layer 152 and the output layer 153 Each of the weight coefficients w ^y _ij and w ^o _jk between the nodes is averaged with Q time-series data and updated.

That is, the higher-order time series prediction generator 62 uses the weighting coefficients w ^h _ij (s + 1), w ^h _jr (s + 1), between the nodes of the input layer 151 and the intermediate layer 152 by the equations (21) to (30). And w ^h _jk (s + 1), and weighting factors w ^y _ij (s + 1) and w ^o _jk (s + 1) between the nodes of the intermediate layer 152 and the output layer 153 are obtained.

Here, η ₂ represents a learning coefficient, and α ₂ represents an inertia coefficient. In Expression (21), Expression (23), Expression (25), Expression (27), and Expression (29), _Δw ^h _ij (s), Δw ^h _jr (s) when s = 1. , Δw ^h _jk (s), Δw ^y _ij (s), and Δw ^o _jk (s) are set to zero.

  In step S36, the upper time series prediction generator 62 determines whether or not the variable s is less than or equal to a predetermined number of learning times. The predetermined number of learning times set here is the number of learning times that the learning error is recognized to be sufficiently small.

  If it is determined in step S36 that the variable s is less than or equal to the predetermined number of learning times, that is, if learning has not yet been performed for the number of times that the learning error is recognized to be sufficiently small, the upper time series is determined in step S37. The prediction generator 62 increments the variable s by 1, and returns the process to step S34. Thereby, the process of step S34 thru | or S36 is repeated. On the other hand, if it is determined in step S36 that the variable s is greater than the predetermined number of learning times, the learning process ends.

  In step S36, in addition to determining the end of the process based on the number of learnings, the end of the process may be determined based on whether or not the learning error is within a predetermined reference value.

As described above, in the learning process of the high-order time series prediction generator 62, the weight coefficients w ^h _ij , w ^h _jr , w ^h _jk , w ^y _ij , and w ^o _jk are obtained for each action sequence, and their averages are obtained. By repeatedly executing the process for obtaining the value, the weight coefficients w ^h _ij , w ^h _jr , w ^h _jk , w ^y _ij , and w ^o _{jk of the} CTRNN 81 used in the generation process are obtained.

Incidentally, in the above-described learning process, the weight coefficient w ^h _ij of each behavior sequence, w ^h _jr, w ^h _jk, was w ^y _ij, and w ^o mean the seek processing _jk to be executed each time the The process may be executed every predetermined number of times. For example, when the predetermined number of learnings to end the learning process is 10,000, the weighting factors w ^h _ij , w ^h _jr , w ^h _jk , w ^y _ij , and w for each action sequence every 10 learning times. ^o Processing for _obtaining the average value of _jk may be executed.

  Next, referring to the flowchart of FIG. 8, the information of FIG. 3 including RNNs 71-1 to 71-N and CTRNN 81 in which the weighting factors obtained by the learning process described with reference to FIGS. 6 and 7 are set. A generation process for generating time-series data by the processing device 51 will be described.

  First, in step S51, the CTRNN 81 of the upper time series prediction generator 62 reads the initial value of the input data. The initial value of the input data here is an initial value supplied to the input node 160 and the context input node 162, for example, a random predetermined value is supplied thereto.

  In step S52, the CTRNN 81 of the high-order time-series prediction generator 62 reads a task ID that identifies an action sequence. The read task ID is supplied to the parameter node 161.

In step S53, the CTRNN 81 of the higher-order time-series prediction generator 62 performs a process of generating the estimated prediction errors errorPredH [N] of the RNNs 71-1 to 71-N at a predetermined time. Details of this generation processing will be described later with reference to FIG. 9, but the CTRNN 81 generates, for example, an estimated prediction error errorPredH ^{t + 1} [N] at time t + 1, and outputs it to the gate signal conversion unit 63.

In step S54, the gate signal conversion unit 63 converts the supplied estimated prediction error errorPredH ^{t + 1} [N] into the gate signal gate ^{t + 1} [N] according to the above-described equation (2), and the converted result is obtained. Output to the gates 72-1 to 72-N.

In step S55, the sensor motor signal sm (t) at time t is input to the RNN 71-n of the lower time series prediction generator 61, and the RNN 71-n corresponds to the input sensor motor signal sm (t) at time t. Thus, the sensor motor signal sm _n (t + 1) at time t + 1 is generated and output to the gate 72-n.

In step S56, the gate 72-n outputs the sensor motor signal sm _n (t + 1) corresponding to the gate signal g ^{t + 1} _n in the gate signal gate ^{t + 1} [N] supplied from the gate signal conversion unit 63. Output. That is, in the gate 72-n, when the gate signal g ^{t + 1} _n is large, the gate is opened widely, and when the gate signal g ^{t + 1} _n is small, the gate is opened small. A sensor motor sm _n (t + 1) corresponding to the degree of opening of the gate 72-n is supplied to the synthesis circuit 73.

  In step S57, the synthesizing circuit 73 synthesizes the outputs from the gates 72-1 to 72-N according to equation (1), and outputs the result of the synthesis as a sensor motor signal sm (t + 1) at time t + 1.

In step S58, the information processing apparatus 51 determines whether to end the generation of time series data. If it is determined in step S58 that generation of time-series data is not terminated, the process returns to step S53, and the subsequent processes are repeated. As a result, the upper time series prediction generator 62 generates an estimated prediction error errorPredH ^{t + 2} [N] at the time t + 2 next to the time t + 1 processed in the previous step S53, and the lower time series prediction generator 61 A sensor motor sm (t + 2) corresponding to the sensor motor signal sm (t + 1) at time t + 1 is generated.

  On the other hand, if it is determined in step S58 that generation of time-series data is to be terminated, for example, when a predetermined number of time steps has been reached, the generation process ends.

Next, the generation process of the estimated prediction error errorPredH [N] in step S53 of FIG. 8 will be described with reference to the flowchart of FIG. FIG. 9 illustrates an example of generating the estimated prediction error errorPredH ^{t + 1} [N] at time t + 1.

First, in step S71, the input node 161-i calculates data x _i (t) according to the equation (3), and the parameter node 161-r calculates data tsdata _r (t) according to the equation (4). , The context input node 162-k calculates the data c _k (t) according to the equation (5) and outputs them.

In step S72, the hidden node 163-j obtains the data h ^u _j (t) by calculating the expression (6), and calculates and outputs the data h _j (t) by the expression (7).

In step S73, the output node 164-i, with the data y ^u _i (t) by calculating equation (8), and outputs data y _i (t) is calculated by Equation (9).

In step S74, the context output nodes 165-k, with the data o ^u _k (t) by calculating equation (10), and outputs the data o _k a (t) calculated by Equation (11).

In step S75, the calculation unit 154 calculates the difference △ x ^u _i a (t + 1) Equation (12), the time t + 1 of the data x ^u _i a (t + 1) calculated by the equation (13), the gate signal converter unit 63 Output.

In step S76, the operation unit 155 obtains the difference Δc ^u _k (t + 1) by the equation (14) and calculates the data c ^u _k (t + 1) at the time t + 1 by the equation (15). Further, the arithmetic unit 155 feeds back the data c ^u _k (t + 1) at time t + 1 obtained as a result of the calculation according to Expression (15) to the context input node 162-k.

In step S77, the arithmetic unit 154 feeds back the data x ^u _i (t + 1) at time t + 1 obtained as a result of the calculation according to Expression (13) to the input node 161-i. And a process returns to step S53 of FIG. 8, and progresses to step S54.

As described above, according to the generation process of FIG. 8, the upper time series prediction generator 62 is an estimation that is an estimation value of a prediction error generated by the RNNs 71-1 to 71-N of the lower time series generator 61 at the time of generation. Prediction error errorPredH [N] is output, and this estimated prediction error errorPredH [N] is converted into a gate signal gate [N] that controls the open / closed states of the gates 72-1 to 72-N. Then, the output signals sm ₁ (t + 1) to sm _N (output from the RNNs 71-1 to 71-N output from the gates 72-1 to 72-N whose open / close states are controlled, represented by the above-described equation (1). The sum of t + 1) is supplied to the sensor and motor of the humanoid robot as a sensor motor signal sm (t + 1) at time t + 1, and the action sequence specified by the task ID is executed.

  Next, additional learning will be described in which the information processing apparatus 51 learns by adding action sequences other than the action sequences A, B, and C learned so far. In the following, the action sequence is such that the robot in the home position grabs the object with both hands, lifts it up to a predetermined height, puts it on the table one level higher than the table on which the object was originally placed, and returns to the home position Learn additional D.

  As described above, different behavioral components are learned in the RNNs 71-1 to 71-N of the lower time series prediction generator 61, respectively. In general, N, which is the number of RNNs 71, is prepared to be sufficiently larger than the number of action parts, and therefore, among the RNNs 71-1 to 71-N, RNNs 71 for which no action parts have been learned (hereinafter referred to as RNN 71). (Also referred to as unused RNN 71 as appropriate).

  In addition to the previously learned behavior sequences A, B, and C, when learning a new behavior sequence D, the RNN 71 that has already learned the behavior component is left as it is and added to the unused RNN 71 It is efficient to learn a new behavior part included in the behavior sequence D. In this case, the RNN 71 learned so far by the learning of the additional behavior sequence D is not broken (the weighting coefficient of the RNN 71 is changed), and the behavior component learned so far is included in the new behavior sequence D. In such a case, the action parts can be used in common.

  Therefore, when the lower time-series prediction generator 61 additionally learns the behavior sequence D, the RNN 71 in which the behavior component has already been learned is given resistance that makes it difficult to change the weighting factor.

The RNN 71 for which the behavioral part has already been learned is the RNN 71-n having a high usage frequency FREQ _n stored in the memory 75 by the process of step S6 in FIG.

Accordingly, the control circuit 76 of the lower time series prediction generator 61, as shown in FIG. 10, it is easy to update the use frequency FREQ _n is less RNN71-n as the weighting factor, RNN71-n use frequency FREQ _n is large, It is difficult to update the weight coefficient, in other words, the learning weight μ _n is determined by the function h ₁ having a negative correlation with the use frequency FREQ _n . The curve represented by the function h ₁ shown in FIG. 10 is a curve having a larger slope as the usage frequency FREQ _n is smaller, and a smaller slope as the usage frequency FREQ _n is larger. In FIG. 10, the function h ₁ is shown as a non-linear curve, but may be a linear straight line as long as it is a function having a negative correlation.

  The additional learning process of the information processing apparatus 51 will be described with reference to the flowchart of FIG.

_First , in step S 101, the control circuit 76 of the lower time series prediction generator 61 reads the usage frequencies FREQ _{1 to} FREQ _N of the RNNs 71-1 to 71-N stored in the memory 75.

In step S102, the control circuit 76 of the low-order time-series prediction generator 61 determines a learning weight μ _n according to the usage frequency FREQ _n of the RNN 71-n using the function h ₁ shown in FIG. The determined learning weight μ _n is supplied to the RNN 71-n.

In step S103, the information processing apparatus 51 learns the time series data of the sensor motor signal corresponding to the action sequence D, the learning process of the lower time series prediction generator 61 in FIG. 6, that is, the processes of steps S1 to S7. Execute. However, in step S5 of FIG. 6 in the process of step S103, the following equation (31) including the learning weight μ _n is employed instead of equation (18).

  After the process of step S <b> 103, the time series data errorL [N] of the prediction error of the action sequence D is stored in the memory 75.

  In step S104, the information processing apparatus 51 reads the time series data errorL [N] of the prediction error of the action sequence D added to the action sequences A, B, and C from the memory 75, and calculates the four prediction errors. For the time series data, the learning process of the upper time series prediction generator 62 of FIG. 7, that is, the processes of steps S31 to S37 are executed. Then, the additional learning process ends.

As described above, in the additional learning process of the information processing device 51, the RNN 71- _n is given the learning weight μ _n that makes it difficult to change the weighting coefficient for the RNN 71-n having a large use frequency FREQ _{n in} the learning so far. Learn -n weighting factors. Thereby, it is possible to efficiently learn the added action sequence without changing the weighting coefficient of the RNN 71 learned so far by learning the additional action sequence D as much as possible.

  Next, another configuration example of the information processing apparatus to which the present invention is applied will be described.

  FIG. 12 shows another configuration example of the information processing apparatus 51. 12, parts corresponding to those of the information processing apparatus 51 in FIG. 3 are denoted by the same reference numerals, and description thereof is omitted.

  The information processing apparatus 51 in FIG. 12 is configured in the same manner as the information processing apparatus 51 in FIG. 3 except that a time filter unit 201 and a nonlinear filter unit 202 are newly provided.

  The time filter unit 201 receives time series data errorL [N] of prediction errors output from the lower time series prediction generator 61. The time filter unit 201 and the non-linear filter unit 202 perform predetermined filter processing on the time series data input thereto, and output the processed time series data to the subsequent stage. The nonlinear filter unit 202 supplies the processed time series data to the upper time series prediction generator 62 as the time series data errorL ′ [N] of the prediction error.

  The high-order time-series prediction generator 62 learns time-series data of prediction errors, but only needs to know rough fluctuations in the prediction errors of the RNNs 71-1 to 71-N in a somewhat long time step. Such fluctuations are not so relevant.

  The time filter unit 201 performs time filter processing on the prediction error time series data errorL [N] output from the lower time series prediction generator 61. That is, the time filter unit 201 performs a so-called low-pass filter process on the time series data errorL [N] of the prediction error output from the lower time series prediction generator 61 and supplies the processed time series data to the nonlinear filter unit 202. To do. For example, as the low-pass filter process, a moving average of a predetermined number of time steps can be used. Thereby, the time series data of the prediction errors of the RNNs 71-1 to 71-N in which minute fluctuations in a short time are suppressed can be supplied to the upper time series prediction generator 62.

  In order for the upper time-series prediction generator 62 to learn rough fluctuations in the prediction errors of the RNNs 71-1 to 71-N in a somewhat long time step, the CTRNN 81 of the upper time-series prediction generator 62 is time-series data. This can also be realized by making the sampling rate when sampling the higher than the sampling rate of the RNN 71 of the lower time series prediction generator 61. For example, the upper time series prediction generator 62 learns time series data obtained by thinning out the time series data of the RNN 71 of the lower time series prediction generator 61 at a predetermined time interval, thereby predicting the RNNs 71-1 to 71-N. A rough variation of the error can be learned. Further, the time sampling can be adjusted by adjusting the coefficient τ in the equations (13) and (15). In this case, the larger the coefficient τ is, the more the variation in the prediction errors of the RNNs 71-1 to 71-N can be learned.

The non-linear filter unit 202 has a function represented by a non-linear curve as shown in FIG. 13 in which the slope is large in the range where the input prediction error errorL _n is small, and the slope is small as the input prediction error errorL _n is large. The input prediction error errorL _n is converted by h ₂ . The nonlinear filter unit 202 supplies the prediction error errorL ′ [N] after the conversion process to the upper time series prediction generator 62.

In the generation processing of the information processing apparatus 51, as described with reference to FIG. 8, the control is performed so that the gate is opened wider as the RNN 72-n has a smaller estimated prediction error errorPredH _n obtained by learning the prediction error errorL [N]. Is done. Conversely, the sensor motor signal sm _n (t + 1) output by the RNN 72- _n having a large estimated prediction error errorPredH _n is hardly used.

Therefore, it can be said that the RNN 72- _n having a smaller estimated prediction error errorPredH _n has a higher contribution rate to the sensor motor signal sm (t + 1) output from the lower time series prediction generator 61 and is more important.

For example, when the prediction error errorL ₁ of the RNN 72-1 and the prediction error errorL _n of the RNN 72- _n are antagonized by a small value between 0 and 1 (for example, 0.3), and between 0 and 1 great value (e.g., 0.9, etc.) considering a case where not conflict with, antagonistic small value between the prediction error ErrorL _n of the prediction error ErrorL ₁ and RNN72-n of RNN72-1 is 0 to 1 In this case, the sensor motor signal sm (t + 1) output from the lower time series prediction generator 61 of the sensor motor signal sm ₁ (t + 1) or sm _n (t + 1) output from the RNN 72-1 or RNN 72-n at the time of generation. ) Is high, it becomes important which of the sensor motor signals of RNN 72-1 and RNN 72-n is superior.

On the other hand, when the prediction error errorL ₁ of the RNN 72-1 and the prediction error errorL _n of the RNN 72-n compete with each other with a large value between 0 and 1, a smaller prediction error is obtained in addition to the RNN 72-1 and the RNN 72-n. The sensor motor signal sm ₁ (t + 1) or sm _n (t + 1) output from the RNN 72-1 or RNN 72 -n at the time of generation is the sensor motor signal output from the lower time series prediction generator 61. Since the rate included in sm (t + 1) is small, it is not so important which of the sensor motor signals of RNN 72-1 and RNN 72-n is dominant.

The non-linear filter unit 202 increases the dominant difference between the RNNs 72 having a small prediction error errorL, which is important for generating the sensor motor signal sm (t + 1), by the function h ₂ and is not important for generating the sensor motor signal sm (t + 1). A process of reducing the dominant difference between the RNNs 72 having a large prediction error errorL is performed. As a result, the upper time series prediction generator 62 can efficiently learn the prediction error errorL output by the RNN 71 important for learning.

  The operations of the time filter unit 201 and the non-linear filter unit 202 are performed as follows. Time series data of Q prediction errors is used as the superordinate time series prediction generator 62 in step S31 of the flowchart described with reference to FIG. When errorL [N] is read from the memory 75 of the lower time series prediction generator 61, time series data errorL ′ [N] of Q prediction errors after being processed by the time filter unit 201 and the nonlinear filter unit 202 are obtained. Read operation.

  Note that both the time filter unit 201 and the nonlinear filter 202 are not necessarily provided at the same time, and only one of them may be provided.

  By the way, in the information processing apparatus 51 shown in FIG. 3 and FIG. 12, as a configuration of the low-order time series generator 61 having a plurality of RNNs 71-1 to 71-n, the outputs of a plurality of RNNs are integrated by a gate mechanism and finally A model called Mixture of RNN Expert that determines the typical output is adopted, but a configuration other than Mixture of RNN Expert can also be adopted.

  As a configuration other than Mixture of RNN Expert, for example, a self-organization map (hereinafter referred to as SOM) used for vector pattern category learning is introduced, and RNN is adopted for each node of SOM. An RNN-SOM that performs RNN parameter learning by selecting an appropriate RNN for external input in a self-organizing manner can be employed. Details of SOM are described in, for example, “T. Kohonen,“ Self-Organizing Map ”, Springer Fairlark Tokyo”, and the like.

  In the Mixture of RNN Expert model shown in FIGS. 3 and 12, all RNNs calculate a learning error (prediction error) for a new learning sample (that is, time series data), and the degree of the learning error. Each RNN learns a learning sample according to

  On the other hand, in RNN-SOM, all RNNs calculate learning errors (prediction errors) for a new learning sample (that is, time-series data). Among them, the RNN with the smallest learning error is calculated. The winner will be determined. After the winner's RNN is determined, the learning error of each RNN does not matter, and the RNN that is close to the winner's RNN learns the learning sample according to the degree of proximity to the winner. The concept of metric space with other RNNs was introduced.

  FIG. 14 is a flowchart of the learning process of the time series data of the sensor motor signal corresponding to the action sequence when the RNN-SOM is adopted as the configuration of the lower time series generator 61.

  The process shown in FIG. 14 is the same as the learning process shown in FIG. 6 except that the process in step S124 is different from the process in step S4 in FIG.

  That is, steps S121 to S123 and S125 to S127 in FIG. 14 are the same as steps S1 to S3 and S5 to S7 in FIG.

In step S124, the low-order time-series prediction generator 61 uses the RNN 71 with the smallest prediction error errorL ^{t + 1} as the winner, and responds to the distance (DISTANCE _n ) from the winner based on the neighborhood function h ₃ shown in FIG. A learning weight γ _n is calculated.

As shown in FIG. 15, the neighborhood function h ₃ is assigned a learning weight γ _{n that} is larger as the RNN 71-n has a shorter distance (DISTANCE _n ) from the winner.

  Next, with reference to FIG. 16 to FIG. 19, an experimental result in which the information processing apparatus 51 described above learns and generates an action sequence performed by the humanoid robot will be described.

  This experiment shows an example of the information processing apparatus 51 of FIG. 12 in which a time filter and a non-linear filter are applied to the time series data errorL [N] of the prediction error output from the lower time series prediction generator 61. Yes. The number N of RNNs 71 in the lower time series generator 61 is 16 (N = 16).

  FIG. 16 shows a result of the information processing apparatus 51 generating the action sequence A after learning the action sequences A, B, and C.

  FIG. 16A shows the output data of the context output node 165 of the CTRNN 141 as the CTRNN 81 of the higher time series prediction generator 62 during the generation process.

  FIG. 16B shows the estimated prediction error errorPredH [N] output from the CTRNN 81 of the higher-order time-series prediction generator 62.

  FIG. 16C shows the gate signal gate [N] obtained by converting the estimated prediction error errorPredH [N] shown in FIG. 16B by the gate signal conversion unit 63.

  16D shows the motor signal of the sensor motor signal sm (t) output from the synthesis circuit 73 of the lower time series prediction generator 61, and FIG. 16E shows the output from the synthesis circuit 73 of the lower time series prediction generator 61. Sensor signals among the sensor motor signals sm (t) thus obtained are shown. In FIG. 16D and FIG. 16E, the data of four motor signals and two sensor signals are shown. However, in order to make the drawing easier to see, a smaller number of data than the actual motor signals and sensor signals are shown. Yes.

  The horizontal axis in FIGS. 16A to 16E represents a time step. 16A, 16D, and 16E represent output values of the context output node 165, the motor signal, and the sensor signal, and are values in the range of 0 to 1. 16B and 16C show the numbers (1 to 16) of the RNN 71 of the lower time series prediction generator 61. FIG.

In FIGS. 16B and C, the value of errorPredH _n or gate signal g ^t _n corresponding to RNN 71-n corresponds to the gray level, and in FIG. 16B, the value of errorPredH _n is small (ie, close to 0). In FIG. 16C, the larger the value of the gate signal g ^t _n (that is, closer to 1), the darker (darker) it is.

  FIG. 17 shows the result of generating the action sequence B by the information processing apparatus 51 after learning the action sequences A, B, and C, and FIG. 18 shows the result of generating the action sequence C.

  FIG. 19 shows the result of the information processing apparatus 51 generating the behavior sequence D after learning the behavior sequence D after learning the behavior sequences A, B, and C.

  17 to 19 is the same except that the illustrated data is related to the action sequences B to D.

  In the generation of time series data corresponding to the action sequence A, as can be seen from FIG. 16C, in the first half of the sequence, the RNN 71-14 is enabled by opening the gate 72-14, and thereafter, in the second half of the sequence. When the gate 72-4 is opened, the RNN 71-4 is activated.

However, conversion to the data shown in Figure 16C from the data shown in FIG. 16B, that is, conversion of the estimated prediction error errorPredH [N] to the gate signal Gate [N] is less the least value among the ErrorPredH ₁ to ErrorPredH ₁₆ Since it is performed using the softmax function of the above formula (2), not the principle of winner-take-all, where a thing is the only winner, it is discrete from a predetermined time (time step) However, the effective RNN 71 is not switched from the RNN 71-14 to the RNN 71-4, but the switching from the RNN 71-14 to the RNN 71-4 is gradually performed over time.

Therefore, even if a plurality of values of errorPredH _{1 to} errorPredH ₁₆ are antagonizing, the winner does not frequently change, and in the antagonizing state, Output can be performed, and thus the learned time-series data can be correctly generated.

  In the generation of the action sequence B, as can be seen from FIG. 17C, the RNN 71-14, the RNN 71-2, the RNN 71-13, the RNN 71-1, and the RNN 71-11 are effective in that order.

  In the generation of the action sequence C, as can be seen from FIG. 18C, the RNN 71-2, the RNN 71-12, and the RNN 71-3 are effective in that order.

  In the generation of the action sequence D, as can be seen from FIG. 19C, the RNN 71-5, the RNN 71-15, the RNN 71-3, and the RNN 71-16 are valid in that order.

  The same applies to the switching of the gate 72 of the action sequences B to D as in the case of the action sequence A of FIG.

That is, the gate signal gate [N] from the RNN 71- _n having the largest estimated prediction error errorPredH _{n at} a predetermined time to the next RNN 71-n ′ (n ≠ n ′) having the largest estimated prediction error errorPredH _{n ′} after a predetermined time. Are switched, the gate signal g _n gradually decreases and the gate signal g _{n ′} gradually increases. That is, the output of the sensor motor signal sm _n (t + 1) is gradually suppressed at the gate 72-n, and the output of the sensor motor signal sm _{n ′} (t + 1) is gradually opened at the gate 72-n ′.

Therefore, even if a plurality of values of errorPredH _{1 to} errorPredH ₁₆ are antagonizing, the winner does not frequently change, and in the antagonizing state, Output can be performed, and thus the learned time-series data can be correctly generated.

  In addition, in the generation result of the action sequence D learned by the additional learning shown in FIG. 19, RNN71-5, RNN71-15, and RNN71-16 that are not valid in the action sequences A to C are valid, It can be seen that the new RNN 71 has learned behavior parts that are not in the behavior sequences A to C learned so far.

  The series of processes described above can be executed by hardware or can be executed by software. When a series of processing is executed by software, a program constituting the software executes various functions by installing a computer incorporated in dedicated hardware or various programs. For example, it is installed from a program recording medium in a general-purpose personal computer or the like.

  FIG. 20 is a block diagram showing an example of the configuration of a personal computer that executes the above-described series of processing by a program. A CPU (Central Processing Unit) 301 executes various processes according to a program stored in a ROM (Read Only Memory) 302 or a storage unit 308. A RAM (Random Access Memory) 303 appropriately stores programs executed by the CPU 301 and data. The CPU 301, ROM 302, and RAM 303 are connected to each other by a bus 304.

  An input / output interface 305 is also connected to the CPU 301 via the bus 304. The input / output interface 305 is connected to an input unit 306 including a keyboard, a mouse, and a microphone, a display including a CRT (Cathode Ray Tube) and an LCD (Liquid Crystal display), and an output unit 307 including a speaker. The CPU 301 executes various processes in response to commands input from the input unit 306. Then, the CPU 301 outputs the processing result to the output unit 307.

  The storage unit 308 connected to the input / output interface 305 includes, for example, a hard disk, and stores programs executed by the CPU 301 and various data. The communication unit 309 communicates with an external device connected directly or via a network such as the Internet or a local area network.

  The drive 310 connected to the input / output interface 305 drives a removable medium 321 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and drives the program or data recorded therein. Get etc. The acquired program and data are transferred to and stored in the storage unit 308 as necessary. Further, the program and data may be acquired via the communication unit 309 and stored in the storage unit 308.

  As shown in FIG. 20, a program recording medium for storing a program that is installed in a computer and is ready to be executed by the computer includes a magnetic disk (including a flexible disk), an optical disk (CD-ROM (Compact Disc-Read Only). Memory), DVD (including Digital Versatile Disc), magneto-optical disk), or removable media 321 which is a package medium made of semiconductor memory, or ROM 302 where the program is temporarily or permanently stored, The storage unit 308 is configured by a hard disk or the like. The program is stored in the program recording medium using a wired or wireless communication medium such as a local area network, the Internet, or digital satellite broadcasting via a communication unit 309 that is an interface such as a router or a modem as necessary. Done.

  In the example described above, the behavior sequences A to C at the time of generation are switched by changing the task ID of CTRNN 81. However, the context input node 162 does not have the task ID input in CTRNN 81. The action sequence A to C at the time of generation may be switched by changing the initial value given to the.

  In this specification, the steps described in the flowcharts include processes that are executed in parallel or individually even if they are not necessarily processed in time series, as well as processes that are executed in time series in the described order. Is also included.

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

It is a figure which shows an example of the conventional information processing apparatus. It is a figure which shows the example of the time series data produced | generated with the information processing apparatus of FIG. It is a figure which shows the structural example of one Embodiment of the information processing apparatus to which this invention is applied. It is a figure which shows the detailed structural example of RNN used for a low-order time series prediction generator. It is a figure which shows the detailed structural example of RNN used for a high-order time series prediction generator. It is a flowchart explaining the learning process of a low-order time series prediction generator. It is a flowchart explaining the learning process of a high-order time series prediction generator. 4 is a flowchart illustrating a generation process of the information processing apparatus in FIG. 3. It is a flowchart explaining the production | generation process in step S53 of FIG. It is a diagram illustrating a function h ₁ that determines the learning weights mu _n in accordance with the use frequency FREQ _n. It is a flowchart explaining the additional learning process of the information processing apparatus of FIG. It is a figure which shows the other structural example of the information processing apparatus to which this invention is applied. Is a diagram illustrating the function h ₂ for converting non-linear depending on the magnitude of the prediction error errorL _n. It is a flowchart explaining the other learning process of a low-order time series prediction generator. It is a diagram illustrating a neighborhood function h ₃ used in the learning process of Figure 14. It is a figure which shows the experimental result of the information processing apparatus 51. FIG. It is a figure which shows the experimental result of the information processing apparatus 51. FIG. It is a figure which shows the experimental result of the information processing apparatus 51. FIG. It is a figure which shows the experimental result of the information processing apparatus 51. FIG. 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

  51 Information Processing Device, 61 Lower Time Series Prediction Generator, 62 Upper Time Series Prediction Generator, 63 Gate Signal Conversion Unit, 71-1 to 71-N RNN, 72-1 to 72-N Gate, 73 Synthesis Circuit, 74 Arithmetic circuit, 75 memory, 76 control circuit, 81 CTRNN, 201 time filter unit, 202 nonlinear filter unit, 301 CPU, 302 ROM, 303 RAM, 308 storage unit

Claims (3)

After learning predetermined time-series data for a plurality of recurrent type neural networks, when adding new time-series data to the plurality of recurrent type neural networks to learn, the predetermined time series data Learning weight determination means for determining a learning weight having a negative correlation with the frequency of use of each of the plurality of recurrent neural networks when the data was previously learned;
Each of the plurality of recurrent neural networks learns the new time-series data according to the learning weight determined by the learning weight determining means. After learning predetermined time-series data for a plurality of recurrent type neural networks, when adding new time-series data to the plurality of recurrent type neural networks to learn, the predetermined time series data Determining a learning weight having a negative correlation with the frequency of use of each of the plurality of recurrent neural networks when the data was previously learned;
An information processing method including a step of learning the new time-series data according to the determined learning weight. After learning predetermined time-series data for a plurality of recurrent type neural networks, when adding new time-series data to the plurality of recurrent type neural networks to learn, the predetermined time series data Determining a learning weight having a negative correlation with the frequency of use of each of the plurality of recurrent neural networks when the data was previously learned;
A program for causing a computer to execute a process including a step of learning the new time-series data according to the determined learning weight.

Images