JPH10326265A - Learning method and device therefor and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To quickly attain learning for a recurrent type neural net. SOLUTION: An initial value is inputted for a recurrent type neural net in a step S21, and a prescribed output is generated. An output is feedbacked to an input for operating rehearsal, and original learning data are recollected in a step S22. New learning data are inputted and learning is operated in a step S23, and the original learning data recollected in the step S22 are inputted and learning is operated in a step S24. The learning with the new learning data and the learning with the original learning data is alternately operated so that learning can be quickly completed.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method, a learning method, and a recording medium, and more particularly, to a learning method, a learning method, a learning method, and a learning method for a recurrent neural network. It relates to a recording medium.

[0002]

2. Description of the Related Art A recurrent neural network uses
You can make predictions. For example, as shown in FIG. 10, the robot 11 moves a route that moves around the obstacle 10A in the counterclockwise direction and a route that moves around the obstacles 10A and 10B in the counterclockwise direction.
Can be stored in a recurrent type neural net included in. If such a memory is stored, for example, when the robot 11 moves around the obstacle 10A,
It stores that landmarks appear in the order of landmark 1, landmark 2, and landmark 5. When moving around the obstacles 10A and 10B, the landmarks are displayed as landmarks. 1, landmark 2, landmark 3, landmark 4, landmark 5
, The robot 11 recognizes these landmarks and recognizes these landmarks.
It can move around A or around the obstacles 10A and 10B.

By the way, for example, if the landmark 4 is deleted from the landmarks 1 to 5 in a state where such learning has already been performed, the recurrent neural network of the robot 11 is Then, it is necessary to perform learning again.

FIG. 11 shows a conventional learning method in such a case. That is, first, step S
At 31, new learning data is input and the new learning data is learned by a recurrent neural network. Step S3
In step 2, the learning result is evaluated, and it is determined whether or not sufficient evaluation has been obtained. If sufficient evaluation has not been obtained, it is determined that learning has not been performed yet, and step S
Returning to 31, the process of inputting new learning data and performing learning is executed again.

[0005] As described above, the learning process is repeatedly executed until it is determined in step S32 that new learning data has been learned.

[0006]

As described above, in the conventional learning method, all the learning processes are restarted from 1 irrespective of a change in which only one landmark is removed. As a result, this is similar to the case where the robot 11 learns from a state in which learning is not performed at all, and there is a problem that learning takes a long time.

[0007] The present invention has been made in view of such a situation, and aims to complete learning more quickly.

[0008]

According to a first aspect of the present invention, there is provided a learning method, comprising: a capturing step of capturing an initial value; and a recalling step of recalling original learning data by rehearsal.
A learning step of adding the new learning data and the original learning data to perform learning.

According to a third aspect of the present invention, there is provided a recording medium for acquiring an initial value, a recalling step for recalling original learning data by rehearsal, and adding new learning data to the original learning data for learning. A program including a learning step is recorded.

According to a fourth aspect of the present invention, there is provided a learning device for capturing an initial value, a recalling unit for recalling original learning data by rehearsal, and learning by adding new learning data and original learning data. Learning means.

In the learning method according to the first aspect, the recording medium according to the third aspect, and the learning apparatus according to the fourth aspect, the original learning data is recalled by rehearsal, and the original learning data and the new learning data are recalled. Learning is performed by adding appropriate learning data.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below. In order to clarify the correspondence between each means of the invention described in the claims and the following embodiments, each means is described. When the features of the present invention are described by adding the corresponding embodiment (however, an example) in parentheses after the parentheses, the result is as follows. However, of course, this description does not mean that each means is limited to those described.

[0013] In the learning apparatus according to the present invention, a fetch means for fetching an initial value (for example, step S21 in FIG. 9).
Recalling means for recalling the original learning data by rehearsal (for example, step S22 in FIG. 9), and learning means for adding the new learning data to the original learning data for learning (for example, step S23 in FIG. 9). , 24).

FIG. 1 shows an external configuration of a robot to which the learning method of the present invention is applied. In this embodiment, a television camera 12 is attached to an upper part of the robot 11 so as to capture a surrounding image. Wheels 13 are attached to the lower side of the robot 11 so that the robot 13 can move to an arbitrary position. A display 14 is attached to a side surface of the robot 11 so that predetermined characters and images are displayed.

FIG. 2 shows an example of the internal configuration of the robot 11. The television camera 12 captures the surrounding video as a color image, and outputs the captured color image data to the control circuit 24 and the quantization circuit 25. The quantization circuit 25 quantizes the input color image data and outputs it to the neural network recognition device 23. The neural network recognition device 23 includes a quantization circuit 25
A landmark described later is recognized from the input color image data, and the recognition result is output to the control circuit 24. For example, the control circuit 24 composed of a microcomputer or the like
3. Notify the moving direction of the robot to 3,
The prediction result of the next landmark supplied from the neural network recognizing device 23 is outputted to the display 14 composed of a CRT, an LCD or the like and displayed.

The control circuit 24 drives the motor 21 to direct the television camera 12 in a predetermined direction. Further, the control circuit 24 drives the motor 22, rotates the wheels 13, and moves the robot 11 to a predetermined position.

FIG. 3 shows the moving space of the robot 11 in a plan view. In this embodiment, the obstacle 10
Landmarks 1 to 5 are arranged around A and the obstacle 10B. In the case of this embodiment, the robot 11 moves counterclockwise around the obstacle 10A on the route of the landmark 1, the landmark 2, and the landmark 5, or moves along the landmark 1, the landmark 2, and the land. Mark 3, Landmark 4, Landmark 5
It is assumed that the vehicle travels in the counterclockwise direction around the obstacles 10A and 10B along the route.

In this case, the control circuit 24 executes the processing shown in FIG. First, in step S1, the control circuit 24 determines whether a landmark has been found.
That is, the television camera 12 captures the surrounding image,
The color image data obtained as a result of imaging is quantized by the quantization circuit 2
5 to the neural network recognition device 23. When the landmark is recognized, the neural network recognition device 23 outputs the recognition result to the control circuit 24 as described later. The control circuit 24 monitors the output of the neural network recognition device 23 to determine whether or not a landmark has been found. If a landmark has not been found yet, the process proceeds to step S2, in which the motor 21 is driven to When the camera 12 is rotated in a predetermined direction,
2, the robot 1 moves the obstacle 10 in FIG.
Wheel 1 so that it moves around A and 10B counterclockwise.
Rotate 3 The processes in steps S1 and S2 are repeatedly executed until it is determined in step S1 that a landmark has been found.

If it is determined in step S1 that a landmark has been found, the process proceeds to step S3, where the control circuit 24 determines whether an obstacle exists in the direction of the found landmark. That is, the control circuit 24
The presence or absence of an obstacle is determined from the output of the television camera 12, and if it is determined that an obstacle exists, the process proceeds to step S
Then, the process proceeds to step 4 in which the wheel 13 is rotated rightward. That is, at this time, the control circuit 24 drives the motor 22 to rotate the wheels 13 clockwise (clockwise).

Thereafter, the flow returns to step S3, and the robot 1
It is determined again whether or not an obstacle exists in the direction in which 1 has newly pointed. If it is determined that an obstacle exists, the process proceeds to step S4 again, and a process of rotating the robot 11 further clockwise is performed. If it is determined in step S3 that there is no obstacle, the process proceeds to step S5, and the control circuit 24 executes a process of moving the robot 11 in the direction of the landmark found in step S1. That is, at this time, the control circuit 24 drives the motor 22, rotates the wheel 13, and moves the robot 11 in the direction of the landmark.

Then, the process proceeds to a step S6, wherein the control circuit 24
Indicates whether or not a landmark has been reached
2 and the output of the neural network recognition device 23. That is, the control circuit 24 receives a signal indicating that a landmark has been detected from the neural network recognition device 23,
Is outputting a sufficiently large landmark image,
It is determined that the landmark has been reached. If the landmark has not been reached yet, the process returns to step S3, and the subsequent processing is repeatedly executed. If it is determined that the landmark has been reached, the process returns to step S1 to find a new landmark. The same processing as that described above is executed toward the landmark.

As described above, the robot 11 travels toward the landmark 1 so as not to collide with the obstacle 10A. When the robot 11 reaches the landmark 1, it travels from the landmark 1 toward the landmark 2. . When the vehicle reaches the landmark 2, the vehicle travels further toward the landmark 5. When the vehicle reaches the landmark 5, the vehicle travels toward the landmark 1 from there.

Alternatively, when the robot 11 reaches the landmark 2,
After moving in the direction of the landmark 3 and reaching the landmark 3, proceed to the landmark 4 from there, and proceed to the landmark 5 when reaching the landmark 4.

The control circuit 24 can program in advance which of the two routes the robot 11 moves.

Here, the configuration of the neural network recognition device 23 will be described. As shown in FIG. 5, the neural network recognizing device 23 includes a correlation storage net 41 composed of a Hopfield neural network, a winner-take-all neural network 42, and a recurrent neural network 43. It is basically configured.

The color image data output from the television camera 12 is input to the quantization circuit 25 and quantized before being input to the correlation storage network 41. That is, as shown in FIG. 6, the quantization circuit 25 has a space (table) composed of hue and saturation, and in the color image data defined by the predetermined hue and saturation, the area A1 Are all red data, for example. Similarly, all the color image data belonging to the area A2 is quantized as green data, and all the color image data belonging to the area A3 is quantized as blue data.

It should be noted that the names of red, green and blue here are merely for convenience, and other names may be used. That is, these names are merely codes (names of quantized data) of each area.

Although the color image data existing in the space defined by the hue and the saturation exists infinitely,
In the case of this embodiment, this is quantized into three pieces of quantized data. In this way, by quantizing a large number of color image data into a sufficiently small number of quantized data,
Learning and recognition of objects by neural networks are possible. As described above, the quantized data obtained by quantizing the color image data by the quantization circuit 25 is supplied to the neural network recognition device 23. Therefore, the quantized data input to the neural network recognition device 23 is data represented by any one of the three data defined by the space shown in FIG.

As shown in FIG. 5, the correlation storage net 41
Has a number of fields corresponding to the number of quantization steps (three in this embodiment) shown in FIG.
The field 41R is a field corresponding to the red quantized data in the area A1 in FIG.
G is a field corresponding to the green quantized data in the area A2 in FIG. 6, and the field 41B is a field corresponding to the blue quantized data in the area A3 in FIG. The input pattern composed of the three pieces of quantized data output from the quantization circuit 25 is stored in the correlation storage net 41.
Are input (recalled) to the neurons in the corresponding fields.

That is, the internal state of each neuron is represented by U
Then, the following equation is established.

[0031]

(Equation 1)

Here, i represents a neuron number, and t
Represents a predetermined time. Thus, U _i ^{t + 1} represents the internal state of the ith neuron at time t + 1.

Here, k is a constant representing a damper,
α is also a predetermined constant.

W _ij represents a connection weight coefficient from the i-th neuron to the j-th neuron. a
_j ^t represents the output of the neuron at the j-th time t. This output is defined by the following equation.

[0035]

(Equation 2)

Here, logistic (A) indicates that A is multiplied by a sigmoid function. T represents a constant. That is, the above equation means that the result of multiplying the result of dividing the internal state of the neuron by the constant T by the sigmoid function is the output of the neuron.

While the recall dynamics are performed as described above, the learning dynamics are represented by the following equations.

[0038]

(Equation 3)

0.5 in the above equation functions as a threshold. That is, the output of each neuron takes a value between 0 and 1, but when the value is smaller than 0.5, the connection weight coefficient is made negative, and when it is larger than 0.5, the function is made that the connection weight coefficient is made positive. have.

When the neural network is made to learn the landmarks 1 to 5 as objects to be recognized, the result of the learning is stored as the connection weight coefficient _Wij .

The winner-take-all type neural network 42 has at least the number of neurons (five neurons in this embodiment) corresponding to the number of landmarks to be recognized. When a predetermined input is made from among the five neurons, the output of one neuron that outputs the largest value is set to 1.0, and the output of the other four neurons is set to 0.0. Is performed. Thereby, one landmark is determined from the pattern defined by the three fields of the correlation storage net 41, and the neuron corresponding to the landmark is fired.

As described above, in the winner-take-all type neural network 42, one neuron fires.
Therefore, the configuration of the recurrent neural network 43 at the subsequent stage for processing the output can be simplified.

The recurrent neural network 43 is basically composed of an input layer 51, an intermediate layer 52, and an output layer 53. The input layer 51 is composed of a pattern node 51A having five neurons corresponding to the winner-take-all neural network 42, a context node 51B having neurons holding the internal state of the recurrent neural network 43, and a control circuit 24. And a direction node 51C having a neuron whose direction of movement is specified.

The output layer 53 has a pattern node 53A having neurons corresponding to five landmarks, and a context node 53B corresponding to the context node 51B in the input layer 51. Each neuron of the intermediate layer 52 connects each node of the input layer 51 and each node of the output layer 53. Also, the context node 5 of the output layer 53
The output of the 3B neuron is fed back to the neuron of the context node 51B of the input layer 51.

When an input corresponding to one landmark is made from the winner-take-all type neural network 42 to the pattern node 51A of the input layer 51, the recurrent neural network 43 predicts a landmark that appears next. Output from the output layer 53.

When color image data is input from the quantization circuit 25, the neural network recognition device 23 executes the processing shown in the flowchart of FIG.

First, in step S11, the process waits until a landmark is searched. In the case of this embodiment, the landmarks 1 to 5 are all colored with a predetermined color, and the neural network recognition device 2
Step S3 is performed when color image data is input.
A determination of YES is made in step 11, and the process proceeds to step S12.

In step S12, the neural network recognizing device 23 executes a process of recognizing the landmark (current landmark) just searched. This recognition processing is performed by the correlation storage network 4 of the neural network recognition device 23.
1 is executed.

When the current landmark recognition process in step S12 is completed, the process proceeds to step S13, where the winner-take-all type neural network 42 narrows down the recognition result obtained in step S12.
That is, it is clear which of the five landmarks has been recognized. Then, the process proceeds to step S14 to perform a process of predicting a landmark appearing next to the current landmark in the recurrent neural network 43. The predicted result is displayed on the display 14. The above processing is repeatedly executed each time a landmark is searched.

Now, landmarks 1 to 5
Is stored (learned) as a connection weight coefficient in the correlation storage net 41. In this state, for example, the correlation storage net 41
Then, when the pattern of the landmark 2 captured by the television camera 12 and quantized by the quantization circuit 25 is input,
In the fields 41R, 41G, and 41B, quantized red data, green data, and blue data of the landmark 2 fire at predetermined positions, respectively. The winner-take-all type neural network 42 determines the corresponding landmark from the firing state of each field,
A neuron corresponding to one landmark is fired based on the determination result. In this case, the neuron corresponding to the landmark 2 fires.

Therefore, the recurrent neural network 4
At the pattern node 51A of the input layer 51 of No. 3, a neuron corresponding to the landmark 2 fires corresponding to the neuron of the winner-take-all type neural network. At this time, the control circuit 24 determines whether the direction to go next is left or right, and inputs a signal corresponding to that direction to the direction node 51C of the input layer 51. In the embodiment of FIG. 5, the neuron corresponding to the left direction is fired. Therefore, the recurrent neural network 43 predicts a landmark arriving next to the landmark 2 and outputs the prediction result to the pattern node 53A of the output layer 53. As shown in FIG. 3, in a state where the landmark 2 is detected, when the next moving direction is the left direction, the landmark that appears next is the landmark 5. Therefore, in this case,
As shown in FIG. 5, in the output layer 53, a neuron of number 5 corresponding to the landmark 5 fires.

When the control circuit 24 receives input of data for predicting the next landmark from the neural network recognition device 23, the control circuit 24 displays a number corresponding to the input data on the display 14.
And display it. In this case, for example, the number 5 is displayed on the display 14. Thereby, the user can know that the landmark that appears next is the landmark 5.

The landmark 2 at the pattern node 51A of the input layer 51 of the recurrent neural network 43
When the neuron corresponding to the right direction is fired by the direction node 51C in the state where the neuron corresponding to is fired, as shown in FIG. Since the mark is the landmark 3, the neuron of the number 3 corresponding to the landmark 3 is fired at the pattern node 53A of the output layer 53.

For example, if the landmark 4 is a landmark having a color similar to that of the landmark 1, it becomes unclear which of the landmark 1 and the landmark 4 has been recognized. However, in the case of this embodiment, since the recurrent neural network 43 is provided with a context node, a state transition is also identified.

That is, as shown in FIG. 3, the landmark 4 appears next to the landmark 3, and the landmark 1 appears next to the landmark 5. In the recurrent neural network 43, since the current state is distinguished by the context nodes 51B and 53B, when the landmark recognized immediately before is the landmark 3,
It is recognized that the landmark to be input next is not the landmark 1 but the landmark 4. Similarly, when the landmark recognized immediately before is the landmark 5, the next predicted landmark is:
It is possible to recognize that the landmark 1 is not the landmark 4.

As described above, the landmark is searched in the recurrent neural network 43 of the robot 11 in the order of the landmark 1, the landmark 2, and the landmark 5, so that the periphery of the obstacle 10A is counterclockwise. And the landmarks are searched in the order of landmark 1, landmark 2, landmark 3, landmark 4, and landmark 5, and the obstacle 10A and the obstacle 1 are searched.
It is assumed that a route that moves around 0B in the counterclockwise direction has already been learned. In this state, for example, it is assumed that the landmark 4 has been deleted as shown in FIG.
Therefore, at this time, the landmarks existing in the work space in which the robot 11 moves are four landmarks of the landmark 1, the landmark 2, the landmark 3, and the landmark 5. When such a slight change is made, the recurrent neural network 4
3 is made to perform learning as shown in the flowchart of FIG.

That is, first, in step S21, a predetermined initial value is input to the recurrent neural network 43. This initial value may be random. In the recurrent neural network 43 to which the initial value has been input, since each neuron has learned the coefficient corresponding to the original learning data, some output is made.

In step S22, the recurrent neural network 43 recalls the original learning data by rehearsal. That is, in the recurrent type neural network 43, the output generated based on the initial value is fed back to the input, and the operation of recalling a new output is repeated based on the fed back output. When such a rehearsal process is performed several times, as described above, since the coefficients of the original learning data have been learned in the neurons of the recurrent neural network 43, the original learning data is added to the recurrent neural network 43. Recall and output data.

Next, the process proceeds to step S23, in which new learning data is input to the recurrent type neural network 43 and learned. That is, a moving route in the order of the landmark 1, the landmark 2, and the landmark 5,
The moving route in the order of the landmark 1, the landmark 2, the landmark 3, and the landmark 5 is learned.

Next, in step S24, the original learning data is input and learned. That is, the moving path in the order of the landmark 1, the landmark 2, and the landmark 5 and the moving path in the order of the landmark 1, the landmark 2, the landmark 3, the landmark 4, and the landmark 5 are learned. As the original learning data, the one recalled by the recurrent neural network 43 by the rehearsal processing in step S22 is used.

Next, the process proceeds to step S25, where it is determined whether or not sufficient learning has been achieved. If it is determined that sufficient learning has not yet been performed, the process returns to step S23,
The subsequent processing is repeatedly executed.

As described above, the learning based on the new learning data and the learning based on the original learning data are added (in the case of this embodiment, alternately arranged) to perform learning. The learning can be completed in a shorter time than in the case where learning is performed using only learning data.

The original learning data may be stored in a memory provided in the robot 11.
However, doing so requires an extra configuration, which not only increases the size of the device but also increases the cost. Therefore, such a method is not very practical.

In the processing example of FIG. 9, the learning using the new learning data and the learning using the original learning data are alternately performed once each. However, for example, the learning is alternately performed twice or three times. It is also possible to make it happen. However, for example, when a total of 3,000 learnings are performed, the learning is first performed 1500 times with the new learning data, and then the next 1500 learnings are performed with the original learning data. A learning result intermediate between the learning result based on the data and the learning result based on the original learning data is obtained, which is not preferable. Therefore, it is desirable to alternate between learning with new learning data and learning with original learning data relatively frequently. Because it is changed relatively frequently, regardless of whether learning with new learning data or learning with the original learning data is performed first,
There is not much difference in the results.

However, for example, after alternately repeating the learning with the new learning data and the learning with the original learning data, the number of times of learning with the new learning data is gradually increased as compared with the learning with the original learning data. You may.

This learning method can be applied not only to the case where the recurrent neural network is applied to the robot but also to the case where it is applied to various devices. However, when applied when the state already learned and the state to be newly learned are relatively similar, a greater effect can be obtained.

[0067]

As described above, according to the learning method according to the first aspect, the recording medium according to the third aspect, and the learning apparatus according to the fourth aspect, the new learning data and the original learning data are obtained. Is added, and learning is performed, so that learning can be completed in a shorter time as compared with the case where learning is performed only with new learning data from the beginning.

[Brief description of the drawings]

FIG. 1 is a diagram showing an external configuration of a robot to which a learning method of the present invention is applied.

FIG. 2 is a block diagram showing an internal configuration of the embodiment of FIG.

FIG. 3 is a diagram illustrating a moving space according to the embodiment of FIG. 1;

FIG. 4 is a flowchart illustrating an operation of the control circuit of FIG. 2;

FIG. 5 is a diagram showing a detailed configuration example of the neural network recognition device 23 in FIG. 2;

FIG. 6 is a diagram for explaining the operation of the quantization circuit 25 of FIG. 2;

7 is a flowchart illustrating the operation of the neural network recognition device 23 of FIG.

FIG. 8 is a diagram illustrating another moving space in the embodiment of FIG. 1;

FIG. 9 is a flowchart illustrating a learning method.

FIG. 10 is a diagram illustrating a space in which a conventional robot moves.

FIG. 11 is a flowchart illustrating a conventional learning method.

[Explanation of symbols]

11 robot, 12 television camera, 13 wheels, 14 display, 23 neural network recognition device, 24 control circuit, 25 quantization circuit,
41 Correlation memory net, 42 Wiener take-all type neural network, 43 Recurrent type neural network

Claims (4)

[Claims] 1. A learning method for a recurrent neural network, comprising the steps of: taking in an initial value; recalling original learning data by rehearsal; adding new learning data to the original learning data; And a learning step of causing the learning method. 2. The method according to claim 1, wherein the addition of the new learning data and the original learning data is performed by alternately performing learning using the new learning data and learning using the original learning data. Item 2. The learning method according to Item 1. 3. A recording medium on which a program for learning a recurrent neural network is recorded, a capturing step for capturing initial values, a recalling step for recalling original learning data by rehearsal, new learning data and original learning. A recording medium characterized by recording a program comprising a learning step of learning by adding data. 4. A learning apparatus for a recurrent neural network, comprising: a fetching means for fetching an initial value; a recalling means for recalling original learning data by rehearsal; and adding new learning data and original learning data to perform learning. A learning device comprising: a learning unit for causing the learning device to perform the learning.