JPH09134338A - Neural network system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable executing reconstruction in accordance with an actual condition in a short time by permitting an adaptive part to change self coupling loads in accordance with the input/output relation of a changed learning object without changing the basic part of a plural layer-neural network. SOLUTION: The neural network(NN) is provided with first to k-th layers, the adaptive part 1 consists of only the first layer and the basic part 2 consists of the second to the k-th layers. Then, when new data is given under conditions where the learned NN control object is changed, only the first layer coupling load wij <(1)> being the NN adaptive part 1 and a threshold value θj <(1)> are permitted to execute relearning so as to be adoptive to the new condition without re-deciding the whole coupling loads of NN and a threshold value and the coupling loads wji <(k)> of the second to the k-th layers and the threshold value θj <(k)> are kept to be fixed. Thus, when the initial condition set at the initial NN learning is changed, NN suitable to the actual condition is reconstructed in a short time by partial correction.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network system that can easily adapt to changes in the controlled object.

[0002]

2. Description of the Related Art When controlling the operation output of a device,
There has been a method using a neural network (hereinafter referred to as NN). As a method of motion control using the NN, for example, first, the forward dynamics of the controlled object is learned by using the NN, then various conditions are given to it, and the error with respect to these conditions is minimized. , There is a method of obtaining an input value that realizes a desired output. As described above, if the learned NN is used, the input pattern to be input is determined with respect to the output pattern to be obtained, and thereby a stable operation output can be obtained. However, if the condition given during the initial learning changes, the controlled object may not operate properly even if the NN is used as it is. The change in conditions means that the device is used in an environment different from the place where the initial learning is performed, or changes while the device to be controlled is in use, for example, the machine of the device itself. Wear and deterioration. When an initial learning condition is changed in this way, an inexpensive NN device having a learning function capable of adapting to an actual environment while maintaining the initial ability, rather than redoing learning of the NN from the beginning, is disclosed in Japanese Patent Laid-Open No. 5-290013. The invention of the NN arithmetic unit is described in. The apparatus of the present invention connects the first arithmetic unit of the NN learned under the initial condition and the second arithmetic unit of the NN that learns the difference caused by the environmental change as the teacher data in parallel, and obtains the arithmetic results of both arithmetic units. That is, the output is obtained by adding them.

[0003]

In the above-mentioned device, the NN of the second arithmetic unit for performing the arithmetic operation on the difference due to the change of the condition has a structure close to the NN of the first arithmetic unit. Therefore, in order to learn the NN of the second calculation unit, all the connection weights must be updated,
It takes a long time to learn and requires a large amount of teacher data. Another NN is required to learn the difference, and the overall configuration is complicated using two NNs, which causes a problem of long calculation time. Then, the subject of this invention is NN
When the initial condition set at the initial learning of is changed, it is possible to reconstruct the NN that meets the actual condition in a short time by partially correcting. Another object is to simplify the structure and shorten the calculation time.

[0004]

According to the present invention, in a system including a neural network having a plurality of layers, an adaptive unit including at least one layer and a basic unit including other layers are provided. It is characterized in that the connection weight of itself is changed and the connection weight of the basic part is not changed so as to match the input / output relationship of the learning target. Second
Of the invention, the error function E between the teacher data and the calculated value,

It is defined by the following equation, and a value obtained by partially differentiating this error E with an input value is obtained by using a backpropagation method through a neural network, and the coupling weight of a specific layer forming the adaptation unit is corrected. The neural network system according to claim 1.

[0005]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The first embodiment shown in FIGS.
The first to Kth layers are provided, and the NN adaptation unit 1 and the NN basic unit 2 are provided.
It is a multi-layer NN consisting of The adaptation unit 1 is composed of only the first layer, and the basic unit 2 is composed of the second to Kth layers. When new data is given under the situation where the controlled object of the NN that has already been learned has changed, all the connection weights w _ji ^{(k) of the} NN and the threshold θ _j ^(k) are redetermined. Instead of, the coupling weight w _{jj of} the first layer which is the NN adaptation unit 1
Only ⁽¹⁾ and the threshold θ _j ⁽¹⁾ are re-learned so as to be adapted to the new condition, and the connection weights w _ji ^{(k) of} the second layer to the Kth layer,
The threshold θ _j ^(k) remains fixed.

Each layer of the NN is constructed as shown in FIG. 2, and the symbols are determined as follows. K: number of layers of NN N _k : number of elements of k-th layer y _j : input to element j of first layer (input layer) (before normalization) (j = 1, ..., N ₁ ) y _j ^{(1 )} : Input to the element j of the first layer (j = 1, ..., N ₁ ) w _jj ⁽¹⁾ : Coupling load from the jth input to the element j of the first layer (j = 1, ..., N) ₁ ) y _j ^(k) : output from the element j of the (k-1 ⁾ th layer (k = 2, ..., K + 1; j = 1, ..., N _k ) w _ji ^(k) : element of the ⁽ k-1 ⁾ th layer The coupling load from i to the element j of the kth layer, (k = 2, ..., K; i = 1, ..., N _k−1 ; j = 1, ..., N _k ) θ _j ^(k) : kth , K; j = 1, ..., N _k ) u _j ^(k) : film potential (k = 1, ..., K ⁾ of the element j of the k-th layer ; j = 1, ..., N k) z j: output from element j of NN (after inverse normalization) (j = 1, ..., N K) a Ij: the normalization factor for the j-th input Number portion (j = 1, ..., N 1) b Ij: deviation of the normalization factor for the j-th input (j = 1, ..., N 1) a Oj: divisor unit normalization factor for the j-th output (j = 1, ..., N K) b Oj: deviation of the normalization factor for the j-th output (j = 1, ..., N K)

At this time, the first layer has only the functions of branching and normalization, and y _j ⁽¹⁾ = (y _j -b _Ij ) / a _Ij (j = 1, ..., N ₁ ) ... (1) u _j ⁽¹⁾ = w _jj ⁽¹⁾ y _j ⁽¹⁾ −θ _j ⁽¹⁾ (j = 1, ..., N ₁ ) (2) y ₁ ⁽²⁾ = u _j ⁽¹⁾ (J = 1, ..., N ₁ ) (3), and for the second to Kth layers, equation (4),
(5).

(Equation 1)

(Equation 2) Also, as in (6), the data is output after being normalized and denormalized. z _j = a _Oj y _j ^{(K + 1)} + b _Oj (j = 1, ..., N _K ) ... (6) Here, the small character {} represents the subscript of the small character. In addition, in the NN of the present embodiment, the coupling load w _{ji of} the first layer
⁽¹⁾ and the threshold value θ _j ⁽¹⁾ are: w _jj ⁽¹⁾ = 1 (j = 1, ..., N ₁ ) ... (7) θ _j = 0 (j = 1, ..., N ₁ ). It is fixed like (8).

At this time, y _jc is the input to the element j of the c-th pattern to be learned (j = 1, ..., N ₁ ) C is the total number of patterns to be learned, and the error function is given by equation (9). Define to.

[Equation 3] Then, to make this E as small as possible,
The coupling weight w _jj ^{(1) of} the first layer, which is the NN adaptation unit 1, and the threshold value θ _j ⁽¹⁾ are changed as in equations (10) and (11).

(Equation 4)

(Equation 5) Here, ε is a sufficiently small positive number, and by properly selecting this ε, the value converges to the minimum value of E. Since the formulas (10) and (11) are basically the same in derivation method, formula (10) will be shown.

Assuming that the subscript c represents the state quantity corresponding to the c-th pattern, the differential formula of the composite function gives

(Equation 6) Can be expanded, and regarding the second term and the third term,

(Equation 7)

(Equation 8) However, the first term cannot be easily obtained. Therefore,

(Equation 9)

(Equation 10) In other words, by propagating the error E from the output to the input direction,

[Equation 11]

(Equation 12)

(Equation 13)

[Equation 14] Decrease k by 1 like

(Equation 15) Is required. At this time, equation (12) is

(Equation 16) Therefore, using the equation (10), w _jj ⁽¹⁾ ,
That is, the coupling load of the first layer is obtained.

For the threshold θ _j ⁽¹⁾ ,

[Equation 17] And the second and third terms are

(Equation 18)

[Equation 19] Becomes The first term is obtained by substituting the equation (18) into the equation (23),

(Equation 20) Therefore, if Expression (11) is used, θ _j ⁽¹⁾ can be sequentially obtained. As described above, if the connection weight and threshold value of the NN basic unit 2 are left as they are and only the connection weight and threshold value of the first layer are obtained, the NN adapted to the new data is obtained.
Re-learning of is completed in a short time.

In the second embodiment shown in FIG. 3, the first layer serving as the adaptation section 1 has a structure different from that of the first embodiment. In the first embodiment, since each neuron in the first layer has a function of inputting only one input signal and branching it, the coupling weight w _jj ⁽¹⁾ becomes the input gain, but In the embodiment, the interference gains from other input signals are also included.
At this time, the formula (12) is

(Equation 21) However, the equation (23) is the same. Then, as in the first embodiment, the connection weight w _ji ^{(1) of} the first layer, which is the adaptation unit 1, and the threshold value θ _j ⁽¹⁾ can be obtained. Since the interference gains from a plurality of incoming signals can be updated in the second embodiment, it is possible to cope with changes in the basic model of the controlled object as compared with the NN of the first embodiment.

The third embodiment shown in FIG. 4 is the same as the first embodiment except that the adaptation section 1 is arranged on the output side instead of the input side. The fourth embodiment shown in FIG. 5 is an NN in which the adaptation unit 1 is formed in the intermediate layer between the basic unit 2 and the basic unit 3. Assuming that the k-th layer configures the adaptation unit 1, equation (12)
Is

(Equation 22) Equation (23) is changed to

(Equation 23) Is changed to. NN as in the third and fourth embodiments
Does not change the input layer, but the output layer and intermediate layer coupling weights and threshold values. Especially, re-learning becomes easy. In addition, in any of the embodiments, the adaptation unit 1 can be configured not only by one layer but also by a plurality of layers. Further, the layers do not necessarily have to be adjacent to each other. The larger the number of layers forming the adaptation unit 1, the wider the adaptation range with respect to the change of the controlled object, but the longer the re-learning time. When the factor of the target change such as the change over the years can be predicted, the re-learning becomes easier by configuring the adaptation unit 1 accordingly.

[0013]

According to the NN of the present invention, when the initial condition set at the time of initial learning is changed, it is possible to reconstruct the NN that matches the actual condition in a short time by partially correcting it. Became. Moreover, since the configuration is simple, the calculation time is shortened.

[Brief description of the drawings]

FIG. 1 is an overall configuration diagram of a first embodiment.

FIG. 2 is a configuration diagram of an NN according to the first embodiment.

FIG. 3 is a NN block diagram of a second embodiment.

FIG. 4 is an overall configuration diagram of a third embodiment.

FIG. 5 is an overall configuration diagram of a fourth embodiment.

[Explanation of symbols]

 1 Adaptation part 2, 3 Basic part

Claims (2)

[Claims] 1. A system comprising a neural network of a plurality of layers, comprising an adaptation section consisting of at least one layer and a basic section consisting of other layers, the adaptation section changing the input / output relationship of a learning target. To match
A neural network system characterized by changing its own connection weight and not changing the connection weight of the basic part. 2. The error function E between the teacher data and the calculated value is given by The value obtained by partially differentiating this error E with an input value is obtained by using a backpropagation method through a neural network, and the coupling weight of a specific layer forming the adaptation unit is corrected. The neural network system according to 1.