JPH11259445A - Learning device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To speedily and precisely execute learning for giving desired output to unknown data. SOLUTION: A plurality of inputs to a learning object are inputted from a signal input means 1 as learning data and an input division means 2 constituted of a plurality of area division means divides the data group into respective areas so as to weight them. A neural network is operated by outputs from the input.division means 2 and a weight updating means 6 obtains the change quantity of the weight coefficient of the neural network so as to update the weight coefficient from error quantity between the output of the neural network and a teacher signal.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning device for learning the characteristics of a target whose parameters are non-linear and unknown from input / output data of the target by a neural network (NN).

[0002]

2. Description of the Related Art A conventional learning device will be described below with reference to the drawings. FIG. 4 shows a conventional schematic diagram. In FIG.
N calculating means, 4 is a teacher signal output means, 5 is an error calculating means, and 6 is a weight updating means.

FIG. 5 is a block diagram of the N.D. FIG. 4 is an explanatory diagram of N calculating means. As shown in FIG. 5, it is a three-layer neural network having m inputs and p outputs, and the number of intermediate layer elements is n
The output function of the hidden layer is a sigmoid function, and the output function of the input / output layer is a linear function. Hereinafter, a conventional example will be described using the three-layer neural network shown in FIG. 5 as an example.

In FIG. 4, the signal input means 1 normalizes a plurality of inputs to a learning target into a range of (0, 1), and N is input to the N operation means 3. N. The N calculating means 3 calculates a neural network output based on the data input from the signal input means.

First, assuming that the inputs to the neural network are X1, X2, X3,..., Xm, the input Uj to the j-th element in the intermediate layer has a threshold value θwj and j from the i-th input layer element. As the coupling coefficient Wij to the th intermediate element,

(Equation 1) Is calculated. Here, W0j = θwj and X0 = 1.

The output Hj of the j-th element in the intermediate layer is

(Equation 2) Becomes

However,

(Equation 3) Is a sigmoid function.

Next, the input Ok to the k-th element in the output layer is
When the threshold is θvk and the coupling coefficient Vjk from the j-th intermediate layer element to the k-th output layer element is:

(Equation 4) Is calculated. Here, it is assumed that V0k = θvk and H0 = 1.

The output Yk of the k-th element in the output layer is

(Equation 5) Is calculated.

Next, in the teacher signal output means 4, the output of the learning target is normalized by (0, 1) and output as a teacher signal. Output Y of N calculating means 3
Sum of the squared error of the teacher signal d and

(Equation 6) Is calculated.

The weight updating means 6 calculates the amount of change of the weight coefficient of the neural network by the back propagation method using the square error sum E as an error function, and calculates the coupling coefficient.

The coupling coefficient Wij from the ith input layer element to the jth intermediate element is

(Equation 7) And the coupling coefficient Vjk from the j-th intermediate element to the k-th output element is

(Equation 8) Is updated. In Equations (7) and (8), μ is a positive constant called a learning rate, α is a positive constant called an acceleration parameter, and ΔW′ij and ΔV′jk are weight coefficient change amounts in previous learning. It is.

By repeating the updating of the weights as described above, the error is reduced, and when the error becomes sufficiently small, the learning is terminated assuming that the output of the neural network has become a desired value.

[0014]

However, in the configuration described above, when the learning target has a complicated nonlinearity, it is necessary to use a large amount of data for learning, and it is desirable for unknown data. Even in the case of learning so as to produce an output, there is a problem that it takes time because it is necessary to perform new learning using previously learned data.

[0015]

In order to solve the above problems, a learning apparatus according to the present invention comprises a signal input means for inputting a plurality of inputs to a learning target as learning data as shown in FIGS. And an input dividing means composed of a plurality of area dividing means for dividing a data set inputted from the signal input means for each area and assigning weights thereto, and calculating a neural network by an output from the input dividing means. Do N. N operation means; teacher signal output means for outputting the output of the learning target as a teacher signal; N calculating means for calculating an error amount between the output of the calculating means and the teacher signal; and the N.N. And a weight updating means for updating the weighting coefficient by obtaining the change amount of the weighting coefficient of the neural network of the N calculating means.

Further, there is provided a learning apparatus comprising an update weight selecting means for selecting a weight coefficient to be updated by the weight updating means according to the size of the data divided for each area by the input dividing means. According to the present invention, even when the learning target has a complicated nonlinearity, it is possible to divide the learning region and perform learning gradually, and also to perform learning on unknown data by using previously learned data. There is no need for new learning, and there is an effect that learning can be performed quickly and accurately.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The invention according to claim 1 of the present invention is a learning apparatus for learning a characteristic of a learning target in a neural network, wherein a plurality of inputs to the learning target are input as learning data. Means, a data set input from the signal input means, an input dividing means comprising a plurality of area dividing means for dividing and weighting the data set for each area, and a neural network operation based on an output from the input dividing means. N. N operation means; teacher signal output means for outputting the output of the learning target as a teacher signal; N calculating means for calculating an error amount between the output of the calculating means and the teacher signal; and the N.N. N is a feature of having a weight updating means for calculating the amount of change of the weight coefficient of the neural network of the arithmetic means and updating the weight coefficient. Even if the learning target has a complicated non-linear system, it is divided into regions. It has the effect of being able to learn gradually and to learn at high speed and with high accuracy. In addition, there is an effect that learning with no accuracy difference can be performed at the boundary between adjacent regions.

According to a second aspect of the present invention, in the first aspect of the present invention, a weight coefficient to be updated by the weight updating means is selected according to the size of the data divided for each area by the input dividing means. It is characterized by having an update weight selecting means to perform learning of a specific area while holding a learning state of another area,
This has the effect of preventing an increase in the number of data and enabling fast and accurate learning.

(Embodiment 1) An embodiment of the present invention will be described below with reference to FIGS. FIG. 1 shows an embodiment of the learning apparatus according to claim 1 of the present invention.
In FIG. 1, 1 is a signal input means, 2 is an input dividing means,
Reference numerals 211 to 2qr denote area dividing means, and 3 denotes N.V. N operation means,
4 is a teacher signal output unit, 5 is an error calculation unit, and 6 is a weight update unit.

In FIG. 1, the signal input means 1 normalizes a plurality of inputs to a learning target into a range of (0, 1) and inputs the result to the input dividing means 2. Here, inputs to the input dividing means 2 are I1, I2,. . . , Iq.

In the input dividing means 2, the input from the signal input means is divided into a plurality of areas by a plurality of area dividing means 211 to 2qr and weighted. FIG.
Shows an example in which the input range of the input I1 is divided into r regions and weighted. As shown in FIG.
, The output I11 passing through the area dividing means 211 to 21r is obtained.
... I1r take values only for the outputs I12 and I13 of the area to which a belongs, and the other outputs I11 and I14 to I1r become zero.

The inputs I2, I3,. . . The region division and weighting of Iq are the same as the input I1. However, the number of regions may be different from each other, or there may be an input that is not divided and weighted. Next, the input dividing means 2
Of the above N. The N operation means 3 performs a neural network operation and calculates an output of the neural network. The number of inputs is m, the number of outputs is p, and the number of hidden elements is n.
Assuming that the output function of the intermediate layer is a sigmoid function and the output function of the input / output layer is a linear function, the neural network output is calculated from (Equation 1) to (Equation 5) as in the conventional example. The teacher signal output means 4 outputs the output of the learning target as a teacher signal, and the error calculation means 5 outputs the N.P. N
The amount of error between the output of the calculating means 3 and the teacher signal is obtained from (Equation 6) in the same manner as in the conventional example. Next, from the output from the input dividing means 2 and the error amount obtained by the error calculating means 5, the N.P. The amount of change of the weighting factor of the neural network of the N calculating means 3 is obtained, and the weighting coefficient is updated by the weight updating means 6 according to (Equation 7) and (Equation 8) as in the conventional example.

The neural network has 4
It may have a configuration of more than layers. Further, the intermediate layer and output layer output functions may be continuous functions other than sigmoid. As described above, even a learning target having a complicated non-linear system can be gradually learned for each region, and can be learned at high speed and with high accuracy. Further, it is possible to perform learning without a difference in accuracy at the boundary between adjacent regions.

(Embodiment 2) FIG. 3 shows a second embodiment of the present invention.
1 shows an embodiment of the learning device of FIG. In FIG. 3, 1 is a signal input means, 2 is an input dividing means, and 211 to 2
qr is the area dividing means, and 3 is the N.N. N calculating means, 4 is a teacher signal output means, 5 is an error calculating means, 6 is a weight updating means, 3
1 is an update weight selection means.

In FIG. 3, the signal input means 1 normalizes a plurality of inputs to a learning target into a range of (0, 1) and inputs the normalized input to the input dividing means 2. In the input division unit 2, the input from the signal input unit is divided into a plurality of regions by a plurality of region division units 211 to 2qr and weighted. Next, the output of the input dividing means 2 is used as an input and the N.P. The N operation means 3 performs a neural network operation and calculates an output of the neural network. Assuming a three-layer neural network having the number of inputs m, the number of outputs p, and the number n of hidden layers, the number of hidden elements is n, the output function of the hidden layer is a sigmoid function, and the output function of the input / output layer is a linear function. (Equation 1) as in claim 1.
The neural network output is calculated from the above (Equation 5). The teacher signal output means 4 outputs the output of the learning target as a teacher signal, and the error calculation means 5 outputs the N.P. The error amount between the output of the N calculating means 3 and the teacher signal is calculated by
Ask from.

Next, the update weight selection means 31 calculates the N.P. From among the weight coefficients of the N calculating means 3, a weight coefficient for updating the weight is selected. FIG. 2
As shown in the above, in the input dividing means 2, the inputs I1, I2,. . . , Iq, and the output after each b division takes one or only two outputs. The above N.P. A weight coefficient on a path in the neural network of the N calculating means 3 is selected.

Next, from the output from the input dividing means 2 and the error amount obtained by the error calculating means 5, the amount of change of the weight coefficient selected by the update weight selecting means 31 is obtained. The weight coefficient is updated by the weight updating means 6 according to (Equation 7) and (Equation 8). The neural network may have a configuration of four or more layers. Further, the intermediate layer and output layer output functions may be continuous functions other than sigmoid.

As described above, since the learning of a specific area can be performed while the learning state of another area is maintained, an increase in the number of data can be prevented, and the learning can be performed at high speed and with high accuracy. it can.

[0029]

As described above, according to the first aspect of the present invention, the input data is divided into regions and weighted.
Even when the learning target has a complicated nonlinearity, learning can be performed gradually by dividing the learning region, and learning can be performed at high speed and with high accuracy. Further, it is possible to perform learning without a difference in accuracy at the boundary between adjacent regions.

According to the second aspect of the present invention, by selecting only the weighting coefficient which has the greatest influence on the learning accuracy of the area to which the input data belongs, and updating the weight, a specific state can be maintained while maintaining the learning state of the other area. Since the learning of the region can be performed, an increase in the number of data can be prevented, and the learning can be performed quickly and accurately.

[Brief description of the drawings]

FIG. 1 is a schematic block diagram of a learning apparatus according to a first embodiment of the present invention.

FIG. 2 is an explanatory diagram of an input dividing unit in Embodiments 1 and 2 of the learning device of the present invention.

FIG. 3 is a schematic block diagram of a learning device according to a second embodiment of the present invention.

FIG. 4 is a schematic block diagram of a conventional learning device.

FIG. 5 is an explanatory diagram of the neural network of FIG. 4;

[Explanation of symbols]

 1 signal input means 2 input division means 211-2qr area division means 3 N calculation means 31 update weight selection means 4 teacher signal output means 5 error calculation means 6 weight update means

Claims (2)

[Claims] 1. A learning device for learning a characteristic of a learning target in a neural network, comprising: signal input means for inputting a plurality of inputs to the learning target as learning data; and a data set input from the signal input means. An input dividing means constituted by a plurality of area dividing means for dividing and weighting each area, and performing a neural network operation based on an output from the input dividing means. N operation means; teacher signal output means for outputting an output of a learning target as a teacher signal; N calculating means for calculating an error amount between the output of the N calculating means and the teacher signal; and the N.N. A learning apparatus comprising: a weight updating unit that obtains a change amount of a weight coefficient of a neural network of an N operation unit and updates the weight coefficient. 2. The method according to claim 1, wherein the size of the data divided for each area by said input dividing means is
2. The learning device according to claim 1, further comprising an update weight selecting unit that selects a weight coefficient to be updated by the weight updating unit.