JP3288066B2 - Learning processing method and system for network configuration data processing device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method and system for a network-structured data processing apparatus such as a neurocomputer for configuring a desired information processing apparatus by learning in accordance with a network-structured data conversion function. , Especially not only from the correct case, but also from counter-examples showing that it should not be,
The present invention relates to a learning processing method and system for a network configuration data processing device that can configure the network configuration data processing device as a desired information processing device.

In recent years, artificial intelligence (AI), factory automation (FA), office automation (O)
With the development of A), there is an increasing demand for intelligent "soft" systems that are easy to use for humans and can coexist with humans. In order to meet this expectation, an adaptive data processing apparatus of a parallel distributed processing method having a network configuration different from that of the conventional Neumann computer has been proposed.

[0003] Among learning methods of internal state values of a data processing device having a hierarchical network structure, particularly, a learning method (DERumelhart, GEHint) called an error back propagation method.
on, and RJWilliams, "Learning Internal Representat
ions by Error Propagation. "PARALLEL DISTRIBUTED
PROCESSING, Vol.1, pp.318-364, The MIT Press, 198
6) is attracting attention because of its practicality.

In a data processing apparatus having a hierarchical network structure, a hierarchical network is composed of nodes of a kind of arithmetic element called a basic unit and an internal connection having a weight corresponding to an internal state value.

FIG. 8 shows the configuration of the basic unit 1. The basic unit 1 has a multi-input, one-output system, and a multiplication unit 2 for multiplying a plurality of inputs by respective internal connection weights, and an accumulation unit for adding all multiplication results thereof. 3 and this accumulated value is subjected to non-linear threshold processing to obtain 1
A threshold processing unit 4 that outputs two final outputs.

As shown in FIG. 9, a number of basic units 1 having such a configuration are hierarchically connected to each other with an input unit 1-h operating by directly distributing an input signal value as an input layer. As a result, a hierarchical network is formed, thereby exhibiting a data processing function of converting an input pattern (input signal) into a corresponding output pattern (output signal).

With respect to this hierarchical network, the error backpropagation method uses a predetermined learning algorithm so that an output pattern from the hierarchical network output for an input pattern prepared for learning becomes a teacher pattern (teacher signal). , The weight value of the internal connection of the hierarchical network is determined. When the weight value is determined by this processing, even if an unexpected input signal is input, a "soft" data processing function that outputs a plausible and appropriate output signal from this hierarchical network is realized. Will be done.

In order to construct a desired information processing apparatus by using a data processing apparatus having a network configuration represented by such a data processing apparatus having a hierarchical network structure, a network structure for defining a data processing function is required. It is necessary to obtain the internal state value by learning processing so that a desired information processing function is realized.

[0009]

2. Description of the Related Art In a data processing device having a hierarchical network structure, if an h layer is a former layer and an i layer is a latter layer,
The accumulation processing section of the basic unit executes the operation of the following equation (1), and the threshold processing section executes the operation of the following equation (2).

[0010] x _pi = Σy _ph w _ih (the sum of Σ is _{h) (1) y pi =} 1 / (1 + exp (-x pi + θ i)) (2) However, h: unit of the h layer number i: i Layer unit number p: Input signal pattern number θ _i : Threshold of the i-th unit of the i-th layer w _ih : Weight of the internal connection of the h-i layer x _pi : From each unit of the h-th layer to the i-th unit of the i-th layer Y _ph : The output from the h-th unit in the h-th layer for the p-th pattern input signal y _pi : The output from the i-th unit in the i-th layer for the p-th pattern input signal The weight value w _ih and the threshold value θ _i are adaptively and automatically adjusted based on error feedback, and learning is performed. As is clear from the equations (1) and (2), the adjustment of the weight value w _ih and the threshold value θ _i needs to be performed simultaneously, but this operation is a difficult operation that interferes with each other. Therefore, the applicant of the present application has filed the application “Japanese Patent Application No. 62-3334
No. 84 (filed on December 28, 1987, "Network Configuration Data Processing Apparatus"), "1" is always output to the h layer on the input side, and a threshold θ _i
The by providing the basic unit 1 allocates a weight value, by incorporating a threshold theta _i in the weight value w _ih, threshold theta _i
It is proposed to treat as a weight value. By doing so, the above-mentioned formula (1) and _{(2), x pi = Σy ph w ih} (Σ is the sum of _{h) (3) y pi =} 1 / (1 + exp (-x pi)) (4 ) Where h: unit number of the h layer i: unit number of the i layer p: pattern number of the input signal w _ih : weight of the internal connection of the hi layer x _pi : number i of the i layer from each unit of the h layer Y _ph : output from the h-th unit in the h-th layer for the p-th pattern input signal y _pi : output from the i-th unit in the i-th layer for the p-th pattern input signal become.

In the error backpropagation method, in order to perform a weight value learning process, an output pattern from an output layer (referred to as a j-layer) and a teacher pattern to be a signal to be taken by the output pattern are firstly calculated according to the following equation. Error vector E _p, which is the sum of squares of the error
The seeking, calculates the total sum E of the error vector E _p as the error of the data processing functions of the hierarchical network.

[0012] _{E = ΣE p = ΣΣ (y} pj -d pj) 2/2 (5) where the sum of the Σ of E _p p, respectively Shigumashiguma the right side p,
j sum E _p: p-th pattern error vector y _pj with respect to the input signal: p th output of the j-th unit of the j layer for the input signal patterns d _pj: j-layer with respect to the input signal of the p-th pattern Teacher signal to the j-th unit Next, assuming that the hierarchical network has a three-layer structure of h layer-i layer-j layer, the sum E of error vectors is asymptotically set to 0 according to the following equation. in order to calculate the update amount [Delta] w _ji and [Delta] w _ih per one weight value.

[0013] _{Δw ji (t) = -ε∂E (} t) / ∂w ji (t) + ζw ji (t-1) (6) Δw ih (t) = -ε∂E (t) / ∂w ih _{(t) + ζw ih (t} -1) (7) where, epsilon: learning constant zeta: momentum t: by repeating the updating of the number of updates the weight value, converges the sum E of the error vector to the minimum value, the _Time weight value w _ji ,
w _ih (threshold _values θ _i and θ _j are also included) is learned.

In the weight value learning process by the error backpropagation method, if the error backpropagation method is applied to a group of prepared learning signals (pairs of input pattern and teacher pattern) for learning, It has been proven that the sum E of the error vectors can be converged to a minimum value. That is, an output pattern corresponding to the teacher pattern can be output.

As described above, in the learning processing method of the weight value of the network structure in the data processing apparatus having the network configuration represented by the error back propagation method, a learning signal for learning is prepared in advance and the weight value is adjusted. Is performed to execute learning.

Therefore, when a desired information processing apparatus is constructed by a data processing apparatus having a network configuration, a sufficient number of learning signals, that is, an input pattern and a teacher pattern meaning a "desirable" output pattern are prepared in advance. In this case, the weight value learning is performed.

[0017]

However, in such a conventional technique which handles only an input pattern and a "desirable" output pattern corresponding to the input signal as a learning signal, that is, only a correct case, for a certain input pattern, There is a problem in that it is not possible to learn an "undesirable" output pattern given in a negative form, that is, no counterexample.

When actually considering information processing in the real world, it is not always possible to define a “desirable” output pattern for a certain input pattern, but rather “undesirable” for a certain input pattern. Defining output patterns can be much easier.

The present invention has been made in view of such circumstances, and a new network for constructing a desired high-performance information processing apparatus by using not only correct examples but also counterexamples as learning signals. An object of the present invention is to provide a learning processing method for a configuration data processing device.

[0020]

FIG. 1 is a diagram illustrating the principle of the present invention. In the figure, reference numeral 10 denotes a network configuration data processing device for constructing a desired information processing device;
Reference numeral 1 denotes a pattern conversion unit provided in the network configuration data processing device 10, which has a network structure and outputs an output pattern corresponding to an input pattern according to a data conversion function defined by the network structure and an internal state value of the structure. Numeral 12 is an internal state value storage unit provided in the network configuration data processing device 10 for managing an internal state value required when the pattern conversion unit 11 executes data conversion processing.

When the pattern converter 11 has a hierarchical network structure, it receives a plurality of inputs and an internal state value (weight value) to be multiplied with these inputs to obtain a product sum, and obtains the product sum. The basic unit 1 that converts the sum of product values by a threshold function to obtain a final output is configured as a basic unit as follows.

An input layer is constituted by a plurality of input units 1-h for receiving and outputting an input pattern, and an intermediate layer provided in one or a plurality of stages is constituted by a plurality of basic units 1-i. An output layer is formed by the basic units 1-j, and between the input units 1-h and the basic units 1-i, between the basic units 1-i, between the basic units 1-i and the basic units 1-j. Are connected to each other, and a hierarchical network structure is realized according to an internal state value set corresponding to each connection.

Reference numeral 20 denotes a learning pattern presenting apparatus which presents a group of input patterns to be learned to the pattern conversion unit 11 and a "desirable" pattern to be taken for each input pattern (this is called an actual teacher pattern). A plurality of types of teacher pattern groups showing “undesirable” patterns (which are referred to as “teacher patterns”) which should not be taken are presented to the internal state value learning device 30 shown below. An internal state value learning device to be mounted, the pattern conversion unit 11 being output in accordance with the presentation of an input pattern by the learning pattern presentation device 20.
An error is defined between the output pattern group from the learning pattern group and a plurality of types of teacher pattern groups such as a real teacher pattern group and a reverse teacher pattern group presented by the learning pattern presentation device 20. An error value is calculated according to the type, and the internal state value is sequentially updated based on the error value.

By the internal state value learning device 30, for example, the error between the output pattern group from the pattern conversion unit 11 and the teacher pattern group is kept as large as possible, while the output pattern group from the pattern conversion unit 11 and the actual teacher pattern group are maintained. By learning the internal state value in which the error value of the network configuration data is within an allowable range, the network configuration data processing device 10 can be set to a desired state not only from a correct case (actual teacher pattern group) but also from a counter example (opposite teacher pattern group). It can be constructed as a high-performance information processing device.

[0025]

According to the present invention, the learning pattern presenting device 20 presents a plurality of types of teacher pattern groups, such as a real teacher pattern and a reverse teacher pattern, to the pattern converter 11, and the internal state value learning device 30 displays the type of the teacher pattern group. In accordance with, detailed internal state value updating and learning are executed.

As described above, according to the present invention, the learning pattern includes not only the correct case group including the set of the input pattern to be learned and the "desirable" output pattern for the learning pattern, but also the input pattern to be learned and the corresponding The network configuration data processing apparatus 10 can be constructed as a desired information processing apparatus from a plurality of types of case groups such as a group of counterexamples including a set of “undesirable” output patterns.

The method of presenting a teacher pattern group (learning curriculum) that the learning pattern presenting device 20 presents to the pattern conversion unit 11 is, for example, first, an actual teacher pattern group is presented, and the output from the pattern conversion unit 11 is output. The internal state value learning device 30 learns an internal state value within an allowable range having an error between the pattern group and the actual teacher pattern group, and then presents a reverse teacher pattern group. While the error with the teacher pattern group is kept large, it is presented stepwise so that the internal state value learning device 30 learns so that the error with the actual teacher pattern group is within a stricter allowable range. By
Since the number of updates of the internal state value of the network configuration data processing device 10, that is, the learning amount can be reduced, the learning efficiency can be improved.

[0028]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below in detail with reference to an example of an information processing device constructed by a data processing device having a hierarchical network configuration.

FIG. 2 shows a system configuration diagram of an embodiment of the present invention. In the figure, reference numeral 60 denotes a data processing device having a hierarchical network configuration having a learning function, which calculates and outputs an output pattern when a certain input pattern is input, and 70 denotes a learning signal management / presentation device. A learning signal (composed of a pair of an input pattern and a teacher pattern including an actual teacher pattern and a teacher pattern) necessary for configuring the hierarchical network configuration data processing device 60 as a desired information processing device is managed. At the same time, the learning signal is presented to the learning function of the hierarchical network configuration data processing device 60.

FIG. 3 shows a detailed configuration example of the hierarchical network configuration data processing device 60. In the figure, reference numeral 61 denotes a hierarchical network unit having a hierarchical network structure as described in FIG. 1, which calculates and outputs an output pattern corresponding to an input pattern according to a data conversion function defined by a weight value; Is a pattern input unit for inputting an input pattern of a learning signal presented from the learning signal management / presentation device 70 to the hierarchical network unit 61 and for inputting a teacher pattern (here, two types of a real teacher pattern and a reverse teacher pattern). ) Is input to a weight value changing unit 66 described later, and 63 is a pattern output unit.
A learning parameter storage unit for outputting an output pattern from the hierarchical network unit 61 to the weight value changing unit 66, a learning parameter storage unit 64 for storing learning parameters ε, ζ, etc. of the backpropagation method, and 65, a weight value storage unit , Which manages the weight value of the internal connection of the hierarchical network unit 61, 66
Is a weight value changing unit, which receives a teacher pattern (actual teacher pattern, on the other hand, a teacher pattern) from the pattern input unit 62 and an output pattern from the pattern output unit 63, and receives the weight value of the internal connection of the hierarchical network unit 61. Are updated and learned so as to approach the target convergence value, and error values for the actual teacher pattern group and the reverse teacher pattern group based on the weight value during learning are output to the learning signal management / presentation device 70.

FIG. 4 shows a detailed configuration example of the learning signal management and presentation device 70. In the figure, reference numeral 72 denotes a learning signal storage unit which discriminates a learning signal (consisting of a pair of an input pattern and a real teacher pattern or a pair of an input pattern and a reverse teacher pattern) between a real teacher pattern and a reverse teacher pattern. What is stored in a form that can be stored, 73 is a learning signal presentation unit,
Reading a learning signal from the learning signal storage unit 72 and presenting it to the hierarchical network configuration data processing device 60;
Is a learning parameter setting unit that sets learning parameters ε and に 対 し て for the hierarchical network configuration data processing device 60, and 75 is a learning convergence determination unit that is performing learning from the hierarchical network configuration data processing device 60. Receiving the error values for the actual teacher pattern group and the reverse teacher pattern group based on the weight values of the two, determines whether or not the error value is within an allowable range, and outputs the determination result to the learning signal presentation unit 7.
3 is notified.

Next, the hierarchical network configuration data processing device 6 executed by the embodiment of the present invention configured as described above.
A detailed description will be given of the contents of processing for constructing a desired information processing apparatus as 0.

Hierarchical network configuration data processing device 60
Is a vector I _p =
_{(I p1, i p2, ...} , i pl) in the output pattern vectors _{O p = (o p1, o} p2, ..., o pm) consisting of m elements assume a case is.

A designer who intends to construct the hierarchical network configuration data processing device 60 as a desired information processing device first creates a teacher pattern group for defining the input / output relationship of the desired information processing device. . As a real teacher pattern group for the input pattern I _yp = (i _yp1 , i _yp2 ,..., I _ypl ), T _yp = (t _yp1 , t _yp2,.
the _ypm), the input pattern _{I np = (i np1, i} np2, ...,
i _npl ), on the other hand, T _np =
(T _np1 , t _np2 ,..., T _npm ) are prepared.

In accordance with the number of elements of the input / output pattern of the hierarchical network configuration data processing device 60, the hierarchical network section 61 of the hierarchical network configuration data processing device 60
Consists of an input layer consisting of the number of input units matching the number of elements in the input pattern, an output layer consisting of the number of basic units matching the number of elements of the output pattern, and an intermediate layer having a practical number of stages and the number of basic units Prepare what is to be done. In the following, it is assumed that the intermediate layer has one stage.

Thus, the hierarchical network unit 61
After completing the preparation, the designer stores the created teacher pattern group in the learning signal storage unit 72 of the learning signal management / presentation device 70.

When the group of teacher patterns is stored in the learning signal management / presentation device 70, the designer who wants the hierarchical network configuration data processing device 60 to function as a desired information processing device is next required. It instructs 70 to start learning the weight value of the hierarchical network unit 61 using the stored teacher pattern group as a learning signal.

Upon receiving this instruction, the learning signal presenting unit 73 of the learning signal managing and presenting device 70 sends the learning signal storing unit 72
, The real teacher pattern group and the reverse teacher pattern group are read one by one, and are processed so as to be presented to the hierarchical network configuration data processing device 60 as a learning signal.

When the presentation of the learning signal is received, the pattern input unit 6 of the hierarchical network configuration data processing device 60
2 inputs the input pattern of the presented learning signal to the hierarchical network unit 61 and also inputs the actual teacher pattern or the reverse teacher pattern to the weight value changing unit 66. Then, the hierarchical network unit 61
By performing data conversion on the input pattern to be input using the weight value of the weight value storage unit 65, a corresponding output pattern is calculated and output to the pattern output unit 63, and the pattern output unit 63 is processed. Performs processing to output this output pattern to the weight value changing unit 66.

As described above, when the real teacher pattern or the reverse teacher pattern is received from the pattern input unit 62 and the output pattern corresponding to the learning input pattern is received from the pattern output unit 63, the weight value changing unit 66
Is calculated for an actual teacher pattern according to the following equation, and the error vector E _yp between the actual teacher pattern and the output pattern corresponding to the input pattern for learning is calculated for each presented actual teacher pattern. calculating the sum of the error vector E _y.

[0041] _{E y = ΣE yp = ΣΣ (} o ypj -t ypj) 2/2 (8) However, the sum of Σ of E _yp p, respectively ΣΣ of the right-hand side p,
j sum E _yp : error vector for input signal of p-th real teacher pattern o _ypj : j for input signal of p-th real teacher pattern
Output from the j-th unit of the layer t _ypj : j for the input signal of the p-th real teacher pattern
The real teacher signal to the j-th unit of the layer. At the same time, for the reverse teacher pattern, the error vector _Enp between the reverse teacher pattern and the output pattern corresponding to the input pattern for learning is calculated and presented according to the following equation. is the calculated total sum of the error vector E _n calculated in contrary every teacher pattern.

[0042] _{E n = ΣE np = ΣΣ {} 1- (o npj -t npj)} 2/2 (9) where the sum of the Σ of E _np p, respectively Shigumashiguma the right side p,
Sum of j E _np : error vector for input signal of p-th _reverse teacher pattern o _npj : output from j-th unit in layer j for input signal of p-th _reverse teacher pattern t _npj : p-th _reverse On the other hand, the inverse teacher signal to the j-th unit in the j-th layer with respect to the input signal of the teacher pattern is expressed by using the error vector as the reciprocal of the distance between the reverse teacher pattern and the output pattern corresponding to the input pattern for learning. , (10), an error vector can be defined.

[0043] E _n = ΣE _np = ΣΣ1 / _{[1+ exp {(o npj -t npj} ) 2 + α} ] (10) where, alpha: constant Then, the weight value changing unit 66, the error E _y, E _n And the update values of the weight values Δw _ji (t) and w _ih (t) according to the following equations:
Is calculated.

[0044] _{Δw ji (t) = k 1} Δw yji (t) + k 2 Δw nji (t) (11) Δw ih (t) = k 1 Δw yih (t) + k 2 Δw nih (t) (12) Δw _{yji (t) = -ε∂E (t} ) / ∂w ji (t) + ζw ji (t-1) (13) Δw yih (t) = -ε∂E (t) / ∂w ih (t) + ζw _{ih (t-1) (14} ) Δw nji (t) = - {- ε '∂E (t) / ∂w ji (t) + ζ' w ji (t-1)} (15) Δw nih (t) = - {- ε '∂E ( t) / ∂w ih (t) + ζ' w ih (t-1)} (16) _{where, Δw yji (t), Δw} yih (t): the real teacher pattern group Update amount of weight value based on Δw _nji (t), Δw _nih (t): update amount of weight value based on teacher pattern group k ₁ , k ₂ : constant ε, ε ': learning constant ζ, ζ': momentum t : according to the updated amount update number of times of the calculation is a weighted value for the next update _{cycle, w ji (t) = w} ji (t-1) + Δw ji (t) (17) w ih (t) = w _{ih (t-1) + Δw} ih (t) (18) the seeking weight value storage section 6 And stores it in.

On the other hand, the learning convergence judging unit 75 of the learning signal management / presentation device 70 outputs the errors E _y , Upon receiving the E _n, the error E _y, by determining whether it is intended to E _n is acceptable, it is determined whether or not the learning has converged, the learning signal presenting unit 73 of the determination result Notice.

Upon receiving the notification of non-convergence from the learning convergence determining unit 75, the learning signal presenting unit 73 executes a process of presenting the same real teacher pattern group and reverse teacher pattern group as learning signals again. Processing will be performed to achieve convergence of learning.

As described above, according to the present invention, the internal state value is determined based on the learning signals of the real teacher pattern group and the reverse teacher pattern group for the network configuration data processing device, thereby enabling the network configuration data processing device to have a desired configuration. It can function as an information processing device.

First, the learning signal management / presentation device 70
Is a phase in which only the actual teacher pattern group is presented as a learning signal, and a phase in which only the teacher pattern group is presented as a learning signal, as a method of presenting the teacher pattern group stored in the learning signal storage section 72. By adopting a method in which the phase of presenting both as a learning signal is switched as necessary and presented, the learning efficiency of the network configuration data processing device 60 can be increased if the learning is advanced. The switching of the phase may be performed automatically based on a predetermined set value, or the learner may observe a display by a simulator or the like having a function of monitoring the learning situation and perform the switching appropriately.

For example, first, only the actual teacher pattern group is used, and the learning convergence judging section 75 sets the allowable range of the error to a gradual one, and proceeds with learning. On the other hand, by using a group of patterns and a group of supervised teachers, by using a method in which the tolerance of the error of the learning convergence determination unit 75 is strict and the learning is further advanced, the learning efficiency of the network configuration data processing device 60 is improved. Can be.

FIG. 5 shows an experimental result according to the present invention. In this experiment, as the hierarchical network unit 61, an input layer consisting of one input unit, an intermediate layer consisting of three basic units,
A three-layer hierarchical network of an output layer composed of one basic unit was prepared.

The learning signals used for the hierarchical network configuration data processing device are indicated by white circles (real teacher pattern group) and black circles (contrary teacher pattern group) in the figure. For example, the circle at the left end in FIG. 5A indicates “output an output signal of 0.25 for an input signal of 0.0”.
Means a real teacher signal, and the black circle on the left side of FIG.
On the other hand, it means that an output signal of 0.2 should not be output for an input signal of 0.375.

The solid line in the figure shows the output result of the hierarchical network configuration data processing device 60 for all inputs after the learning signal is learned by the hierarchical network configuration data processing device 60 by the above-described method. It is. Loop in the figure represents the number of times of learning.

For example, as shown in FIG. 5 (a), as a result of learning using only the same actual teacher pattern group as before, the so-called undershoot and overshoot appear before and after the input signal of 0.5. A distortion equivalent to a shot appeared. Therefore, in order to efficiently correct this part, learning was performed by adding a group of teacher patterns indicated by black circles in FIG. As a result, a desired output as shown by a solid line in FIG.

FIG. 6 shows a state in which another learning pattern is learned by the conventional technique. As shown in the figure, as a result of repeating the learning 4533 times using the real teacher pattern group for the six input patterns, an output as shown by a solid line is obtained for each input.

FIG. 7 shows the results of an experiment in which learning was performed by applying the present invention to the same case as in FIG. First, learning was repeated 1000 times using only the actual teacher pattern group. As a result, input / output characteristics as shown in FIG. 7A were obtained. Therefore, if the output signal corresponding to the input signal at the position of the mark x shown in FIG. 7A is added as a teacher pattern and learning is further continued, the result is obtained as shown in FIG. Became. As is clear from the comparison between the result shown in FIG. 6 and the result shown in FIG. 7B, it can be seen that the internal state value can be quickly converged to a desired value by using the teacher pattern group.

In the description of this embodiment, the case where learning is performed using two types of teacher pattern groups, a real teacher pattern group and a reverse teacher pattern group, has been described as an example. Depending on the degree of undesirability, such as those for which the output pattern group should not take values and those for which the output pattern group should not be possible, the internal state is determined according to the undesired levels. The update amount of the value may be adjusted.

The present invention is not limited to a data processing device having a hierarchical network structure, and can be similarly applied to learning processing of a data processing device having another network structure. Also, it is clear that the learning method of the internal state value of the network can be applied not only to the back propagation method. Further, the method of defining the error between the teacher pattern group and the output pattern group from the network configuration data processing device 60, the method of determining the update amount of the internal state value based on the error, the method of presenting the teacher pattern group, and the like are also described above. The present invention can be implemented without being limited to the method of the embodiment.

[0058]

As described above, according to the present invention,
As a learning pattern, not only a correct case group consisting of a set of an input pattern to be learned and a "desirable" output pattern corresponding thereto, but also a counterexample group consisting of a set of an input pattern to be learned and a "desirable" output pattern corresponding thereto. From the plural types of case groups, the network configuration data processing device can be constructed as a desired high-performance information processing device. In addition, the efficiency of learning can be further improved by changing the presentation method such as a phase using only correct case groups, a phase using only counterexample groups, and a phase using both as the case group presentation method. Can be improved.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a system configuration diagram of an embodiment of the present invention.

FIG. 3 is a diagram illustrating a configuration example of a hierarchical network configuration data processing device illustrated in FIG. 2;

FIG. 4 is a diagram illustrating an example of a configuration of a learning signal management and presentation device illustrated in FIG. 2;

FIG. 5 is an explanatory diagram of an experimental result according to the present invention.

FIG. 6 is an explanatory diagram of an experimental result according to a conventional technique.

FIG. 7 is an explanatory diagram of an experimental result according to the present invention.

FIG. 8 is a basic configuration diagram of a basic unit.

FIG. 9 is a basic configuration diagram of a hierarchical network.

[Explanation of symbols]

 Reference Signs List 10 network configuration data processing device 11 pattern conversion unit 12 internal state value storage unit 20 learning pattern presentation device 30 internal state value learning device 60 hierarchical network configuration data processing device 70 learning signal management presentation device

Continuation of the front page (56) References JP-A-5-334276 (JP, A) JP-A-5-20291 (JP, A) JP-A-5-165801 (JP, A) JP-A-2-170265 (JP) , A) Mitsumi Yamada, et al. “A learning method in which the difference between the teacher signal and output increases”, Proc. Of the 1991 IEICE Autumn Meeting, Volume 6, Japanese Society of Electronics, Information and Communication Engineers, Japan , Japan Patent Office Information Center: November 13, 1991, pp. 12 (58) Field surveyed (Int.Cl. ⁷ , DB name) G06N 1/00-7/00 G06G 7/60 G05B 13/00-13/04 G10L 3/00-9/20 G06K 9/00- 9/82 G06T 7/00-7/60 CSDB (Japan Patent Office) JICST file (JOIS)

Claims (5)

(57) [Claims] 1. A network configuration data processing device for calculating and outputting an output pattern corresponding to an input pattern and having a network structure in which a plurality of operation elements for performing data conversion specified by an internal state value on an input are connected. a learning processing method for determining by learning the internal <br/> unit status value of the location, in response to the input pattern group of the learning object to be presented to the network configuration data processing equipment, a pattern to take the output pattern group A plurality of types of teacher pattern groups including a real teacher pattern group to be instructed and a pattern group for instructing a pattern that the output pattern group should not take are given as learning information, and the network configuration data based on the input pattern group to be learned is provided. an output pattern group of the processing instrumentation placed et the error value between the teacher pattern group, calculated for each type of teacher pattern group The update amount of the internal state value is determined based on a plurality of error values obtained according to the type of the teacher pattern group, and while the internal state value is sequentially updated using the update amount, the value related to the error value is determined. double but to a predetermined allowable range, a learning processing method for Rene Ttowaku configuration data processing device to determine the learning the internal state value, used for learning of the network configuration data processing device
Several types of teacher patterns are processed by the network configuration data processing.
Output pattern group from the logical device and the actual teacher pattern group
According to the error value of the actual teacher pattern group, then other
Of the teacher patterns including the teacher patterns
A learning processing method for a network configuration data processing device, wherein the learning is advanced by sequentially updating internal state values . 2. The learning processing method for a network configuration data processing device according to claim 1, wherein said learning to be presented to said network configuration data processing device is performed.
As a method of presenting the teacher pattern group,
Phase that uses only the training,
Phase using only pharma pattern groups, using both
Has a phase, the learning processing method for network configuration data processing apparatus characterized by learning while switching these phase at any time. 3. A learning processing method for a network configuration data processing device according to claim 1, wherein said network configuration is based on an input pattern group to be learned.
Output patterns from the data processing device and the teacher pattern
The output pattern group and the plurality of types
The distance is defined between each of the teacher pattern groups, and at least the output pattern group and the actual teacher pattern group
To reduce the distance between and the output pattern group
On the other hand, the distance between the teacher pattern group and the
Thus, the internal state of the network configuration data processing device
The values are sequentially updated, and the error value between the output pattern group and the actual teacher pattern group is allowed.
A learning processing method for a network configuration data processing device, wherein learning is performed so as to fall within a range . 4. A learning processing method for a network configuration data processing device according to claim 1, wherein said network configuration is based on a group of input patterns to be learned.
Output patterns from the data processing device and the teacher pattern
The output pattern group and the actual teacher
The distance between the turn group and the above-mentioned teacher pattern group
At least an output pattern group and the actual teacher pattern group
To reduce the distance between and the output pattern group
On the other hand, the distance between the teacher pattern group and the
Thus, the internal state of the network configuration data processing device
The values are sequentially updated, and the distance and the distance between the output pattern group and the actual teacher pattern group are updated .
And the distance between the output pattern group and the reverse pattern group
The learning processing method of the network configuration data processing device, wherein learning is performed so as to be within an allowable range . 5. A network configuration data processing device for calculating and outputting an output pattern corresponding to an input pattern and having a network structure in which a plurality of operation elements for performing data conversion specified by an internal state value on an input are connected. And a learning processing system for a network configuration data processing apparatus that learns the internal state value based on a group of input patterns to be learned and a group of teacher patterns predetermined for the group of input patterns. , and the addition to presenting an input pattern group of the learning object to be input to the network structure data processing equipment, in accordance with their input pattern group, the network configuration data processing equipment
As a teacher pattern group for instructing pressurization et output pattern group, the plurality including a real teacher pattern group that instructs the pattern to take the output pattern group, and although the teacher pattern group for instructing the pattern output pattern group must not take wherein the type of teacher pattern group and the learning pattern presentation means to presentation as learning information, and the network configuration data processing instrumentation placed these output pattern group according to the input pattern group of the learning object given depending on the learning pattern presentation hand stage the error value between the teacher pattern group, calculated for each type of teacher pattern group, based on the plurality of error values obtained in accordance with the type of the teacher pattern group, the internal state of the network configuration data processing equipment While determining the update amount of the value and sequentially updating the internal state value using the update amount, at least the output pattern group and the training A learning processing system of Ne Ttowa <br/> click configuration data processor error value and a internal state value learning hand stage performing learning control such that the allowable range of the teacher pattern group, the learning pattern The presenting means includes the network configuration data.
Of multiple types of teacher patterns used for learning data processor
Is the output pattern from the network configuration data processor.
According to the error value between the pattern group and the actual teacher pattern group,
The actual teacher pattern group, and then the other teacher pattern group
Means for presenting in the order of a group of teacher patterns, wherein said internal state value learning means comprises:
Updates the internal state values sequentially based on the information presented by
A learning processing system for a network configuration data processing device, characterized in that learning is advanced .