JP2744313B2 - Network configuration data processing device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] The present invention relates to a network configuration data processing device that performs adaptive data processing according to a network configuration data conversion function, and aims to realize more accurate data processing. One or more inputs and an internal state value to be multiplied with respect to the inputs to obtain a product sum value, and an internal connection of a plurality of basic units that obtains an output value by performing a function conversion on the product sum value And a network structure having one or more input units for receiving an input signal value of an input pattern and distributing the input signal value to a corresponding basic unit, wherein a network structure / internal state value of the network structure A network that calculates and outputs an output pattern corresponding to an input pattern according to a specified data conversion function In the network configuration data processing device, at least one of the input units divides a space of signal values input to the own unit and allocates an attribute function for defining an attribute value of an input signal value to each of the divided spaces. In addition, when an input signal value is received, attribute values corresponding to the number of the attribute functions are calculated in accordance with the attribute functions and input to the corresponding basic units.

[Industrial applications]

The present invention relates to a network configuration data processing device that performs adaptive data processing in accordance with a network configuration data conversion function, and more particularly to a network configuration data processing device that can realize more accurate data processing.

With a conventional sequential processing computer (Neumann-type computer), the data processing function cannot be adjusted according to changes in the method of use or environment. An adaptive data processing device according to a distributed processing method has been proposed. With this network configuration data processing device, without creating an explicit program,
An output signal (output pattern) from the network structure output in response to the presentation of the input signal (input pattern) prepared for learning is made to match the teacher signal (teacher pattern) according to a predetermined learning algorithm. The weight value of the inner connection is determined. When the weight value is determined by the learning process, a "soft" data processing function of outputting an appropriate output signal from the network structure even if an unexpected input signal is input is provided. Will be realized.

In order to make the network-structured data processing device having such a configuration practical, it is necessary to take measures for realizing higher-precision data conversion processing.

[Conventional technology]

The prior art will be described according to a network configuration data processing device having a hierarchical network configuration.

In a data processing device having a hierarchical network configuration, a hierarchical network is configured from a kind of node called a basic unit and an internal connection having a weight value corresponding to an internal state value. FIG. 17 shows the basic configuration of the basic unit 1. The basic unit 1 is a multi-input, one-output system, and a multiplication unit 2 for multiplying a plurality of inputs by respective weights of internal connections, and an accumulation unit 3 for adding all the multiplication results thereof. And a threshold processing unit 4 that performs a non-linear threshold process on the accumulated value and outputs one final output.

Assuming that the h-th layer is the first layer and the i-th layer is the second layer, the accumulation processing unit 3 executes the operation of the following expression (1), and the threshold processing unit 4 executes the operation of the following expression (2).

y _pi = 1 / (1 + exp (−x _pi + θ _i )) (2) where h: unit number of h layer i: unit number of i layer p: pattern number of input signal θ _i : i number of i layer Unit threshold value W _ih : Weight value of internal connection between _{hi and} i layers x _pi : Sum of products of inputs from each unit of h layer to unit i of y layer y _ph : h layer of input signal of p-th pattern Output from the h-th unit y _pi : Output from the i-th unit in the i-th layer with respect to the p-th pattern input signal In the conventional hierarchical network configuration data processing device, a number of basic units 1 having such a configuration The input unit 1 'which distributes and outputs the signal value as it is is used as an input layer to form a hierarchical network by being hierarchically connected as shown in FIG. 18, and to convert the input pattern into a corresponding output pattern. Demonstrate the parallel data processing function It becomes door.

In a data processing device having a network configuration, it is necessary to obtain a weight value of a network structure that defines a data conversion function by a learning process. As a learning processing method of the weight value of the hierarchical network configuration data processing device, in particular, a learning processing method called a back propagation method (DERumelhart, GE Hinton, and RJ Williams, “Le
arning Internal Representationsby Error Propagatio
n ", PARALLEL DISTRIBUTED PROCESSING, Vol.1, pp.318-3
64, The MIT Press, 1986) has attracted attention because of its practicality.

In the back propagation method, learning is performed by adaptively and automatically adjusting the weight value W _ih and the threshold value θ _i of the hierarchical network by error feedback. (1)
As is apparent from the equation (2), the adjustment of the weight value W _ih and the threshold θ _i needs to be performed at the same time.
It is a difficult task to interfere with each other. Therefore, the applicant of the present application has filed a Japanese Patent Application No. 62-333484 (December 28, 1987)
As disclosed in Japanese Patent Application, “Network Configuration Data Processing Apparatus”), by providing a unit that always outputs “1” to the h layer on the input side and assigns a threshold value θ _i to the output as a weight value. was suggested to treat the threshold theta _i as Kasami value incorporates threshold theta _i in the weight value W _ih. By doing so, the above (1)
Equation (2) is y _pi = 1 / (1 + exp (−x _pi )) (4)

Next, a description will be given of a weight value learning processing method by the back propagation method according to the description format of the expressions (3) and (4). Here, this description will be made on the assumption that the hierarchical network unit has a hierarchical network having a three-layer structure of the n-th layer-the j-th layer shown in FIG.

By analogy with equations (3) and (4), the following equations (5) and (6)
An expression is obtained. That is, y _pi = 1 / (1 + exp (−x _pj )) (6) where j: the unit number of the i-th layer W _ji : the weight of the internal connection between the i-j layers x _pj : the j-th layer from each unit of the i-th layer Y _pj : Output from the j-th unit in the j-th layer for the p-th pattern input signal In the back propagation method, the signal is output when an input pattern for learning is presented When the output pattern y _pj from the output layer and the teacher pattern d _pj (teacher signal to the j-th unit and j-th unit for the input signal of the p-th pattern) which is a signal to be taken by the output pattern y _pj are determined, first, Next, a difference value [d _pj −y _pj ] between the output pattern y _pj and the teacher pattern d _pj is calculated, and then α _pj = y _pj (1−y _pj ) (d _pj −y _py ) is calculated. ,continue, Here, ε: learning constant ζ: momentum t: number of times of learning According to, the update amount ΔW _ji (t) of the weight value between the i-th layer and the j-th layer
Is calculated. Here, for adding those pertaining to the update amount of the weight values determined in the previous update cycle of "ζΔW _ji (t-1)" is order to speed up the learning.

Subsequently, using the calculated α _pj , first, , Then , The update amount ΔW _ih (t) of the weight value between the h layer and the i layer
Is calculated.

Subsequently, the weight value W _ji for according to the update amount calculated as described above in the next update cycle _{(t) = W ji (t} -1) + ΔW ji (t) W ih (t) = W ih (t-1) + ΔW By repeating the method of determining _ih (t), the output pattern y _pj from the output layer output when the input pattern for learning is presented and the signal to be taken by the output pattern y _pj Processing is performed to learn weight values W _ji , W _ih (including threshold values θ _i , θ _j ) at which a certain teacher pattern d _pj matches.

Then, the hierarchical network unit is composed of g layer-h layer-i layer-j
If you have a 4-layer hierarchical network called layers,
First of all, , Then The update amount ΔW _hg (t) of the weight value between the g layer and the h layer is calculated according to the following formula. By determining while using, all weight values and threshold values are processed to be learned.

[Problems to be solved by the invention]

However, when a learning algorithm such as a back propagation method is applied to a network configuration data processing device such as a conventional hierarchical network configuration data processing device, learning of a weight value of a network structure is performed. May not converge. If the learning of the weight value does not converge, a desired output pattern cannot be output with sufficient accuracy even if an input pattern for learning is input. There was a problem that it could not be executed.

As a method for solving such a problem, it is conceivable to adopt a method such as increasing the number of intermediate layers in the hierarchical network structure or increasing the number of units in the intermediate layer. However, if this method is adopted, it is necessary to increase the number of basic units 1. However, since the hardware configuration of the basic unit 1 is quite complicated, this method requires a large amount of hardware. A new problem that would lead to a significant increase in As another method, it is conceivable to adopt a method of increasing the number of units in the input layer. However, this method cannot be applied to the case where the number of input signals is limited, and the learning signals are already prepared. In this case, there is a problem that the learning signal must be collected again.

The present invention has been made in view of such circumstances,
It is an object of the present invention to provide a new network data processing device capable of improving the accuracy of data conversion processing of a network configuration data processing device that performs adaptive data processing.

[Means for solving the problem]

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

In the figure, 10 is a network configuration data processing device equipped with the present invention, which performs adaptive data processing functions with high precision according to the network structure, and 11 is a network structure unit, And an internal state value to be multiplied with respect to the input to obtain a sum-of-products value, and to convert the product-sum value into a function to obtain an output value. And a network structure having one or more input units for receiving an input signal value of an input pattern and distributing the input signal value to a corresponding basic unit, and is defined by the network structure and an internal state value of the structure. 12 calculates and outputs an output pattern corresponding to the input pattern according to the data conversion function, and 12 is an internal state value storage unit. Network structure 11 manages the internal state value required upon execution of the data conversion process.

1'-h is an input unit constituting the network structure unit 11, which receives an input signal value of an input pattern to be input and processes the input signal value so as to distribute the input signal value to a corresponding basic unit of the network structure unit 11. Thing, 1'-
hq (q = 1 to n) is an attribute value calculation unit connected to a specific input unit 1'-h, and an attribute corresponding to an input signal value received by the input unit 1'-h according to the assigned attribute function. A value is calculated, and an attribute calculated in place of the input signal value received by the input unit 1'-h is given to the basic unit of the network structure unit 11 to which the input unit 1'-h is output. It processes to distribute values. In the present invention, at least one input unit 1'-h of the network structure unit 11 is provided.
Are provided with this attribute value calculation unit 1′-hq. Note that the symbol h represents an input layer.

1-i and 1-j are basic units that constitute the network structure unit 11, and receive one or more inputs and an internal state value to be multiplied with the input to obtain a product sum value. , And obtains an output value by performing a function conversion on the product-sum value. Note that the symbol i represents an intermediate layer, and the symbol j represents an output layer.

The network structure unit 11 may take a network structure other than the hierarchical network structure. However, when taking the hierarchical network structure, the input unit 1'-h and the basic unit 1-i, j are used as constituent units, and The input unit 1'-h constitutes an input layer, the plurality of basic units 1-i constitute an intermediate layer provided in one or more stages, and the one or more basic units 1-i
j constitutes an output layer, and between the input unit 1'-h and the basic unit 1-i, the basic unit 1
-I, basic unit 1-i and basic unit 1-
j, and a hierarchical network structure is realized according to an internal state value set corresponding to each connection. Then, according to the hierarchical network structure of the network structure unit 11, the attribute value calculation unit 1'-hq
Indicates the attribute value to be calculated as the basic unit 1 of the first intermediate layer.
-Process to distribute to i.

[Action]

According to the present invention, each of the n attribute value calculation units 1'-hq connected to the same input unit 1'-h executes, for example, an attribute function conversion function as shown in FIG. Thus, the signal value space of the input signal value X input to the input unit 1'-h of the connection source is configured to be divided. The attribute value of this attribute function may be any value as long as it is at a level that can be handled by the network structure unit 11. For example, the attribute function is defined by a membership function of a fuzzy set and the input signal value is defined. When dividing the signal value space of X, the attribute value takes a value between “0” and “1” defined by the membership function.

The input signal input to the input unit 1'-h is characterized by an n-dimensional attribute value according to the processing of the attribute value calculation unit 1'-hq. For example, n
= 3, three attribute value calculation units 1'-hq (q = 1
When attribute functions such as the second view is assigned with respect to 3), each attribute value calculating unit 1'-hq, when the 0 ≦ X ≦ K _1, only the attribute value calculating unit 1'-h1 Outputs an attribute value that is not a zero value according to the X value, and when K ₁ ≦ X ≦ K ₂ , only the attribute value calculation unit 1 ′ -h 1 and the attribute value calculation unit 1 ′ -h 2 If an attribute value that is not a value is output and K ₂ ≦ X ≦ K ₃ , only the attribute value calculation unit 1 ′ -h 2 outputs an attribute value that is not a zero value according to the X value, and K ₃ ≦ X ≦ K ₄ sometimes, only the attribute value calculating unit 1'-h2 and attribute value calculating unit 1'-h3 outputs the attribute value non-zero value corresponding to the X value, when the K ₄ ≦ X ≦ K _5, the attribute value calculating unit 1 '-H3 outputs an attribute value that is not zero according to the X value, and calculates the attribute value as follows. Work structure 11
To the corresponding basic unit 1 (the basic unit 1-i constituting the intermediate layer when the network structure unit 11 has a hierarchical network structure).

The attribute value calculated in this way is, for example, a size having the input signal value X, as described in the case where the attribute function is a membership function expressing “ambiguity” of the magnitude of the input signal value X. It expresses the degree of “ambiguity”.

Thus, in the conventional case, the input unit 1'-
The input signal value X received by h is the same as the corresponding basic unit of the network structure unit 11 (the network structure unit 11).
Is distributed to the basic units 1-i) constituting the intermediate layer when the hierarchical network structure is adopted,
According to the present invention, an attribute value, which is a semantic expression of the input signal value X, is calculated, and the corresponding basic unit 1 of the network structure unit 11 (when the network structure unit 11 has a hierarchical network structure, forms an intermediate layer). To be distributed to the basic units 1-i). For example,
To explain, instead of inputting “red”, data having more meaning such as “white whitish” or “yellowish red” (which can have various meanings by a combination of attribute values) is stored in the network structure unit 11. Are allocated to the corresponding basic unit 1 (the basic unit 1-i constituting the intermediate layer when the network structure unit 11 has a hierarchical network structure). From now on, the network structure part will be
Even when learning of weight values assigned to the 11 units is performed, it is possible to learn a weight value with a smaller error than before. Therefore, by using the present invention, the accuracy of the data conversion function of the network configuration data processing device can be improved.

〔Example〕

Hereinafter, the present invention will be described in detail according to an embodiment applied to a hierarchical network configuration data processing device.

FIG. 3 shows an embodiment of a hierarchical network structure of a hierarchical network configuration data processing apparatus equipped with the present invention. In the drawing, 1'ah is a plurality of input units constituting an input layer of a hierarchical network structure, 1-i is a plurality of basic units constituting an intermediate layer of a hierarchical network structure, and 1-j is a hierarchical network structure. Are a plurality of basic units that constitute the output layer of FIG. These basic units 1
-I, 1-j is configured to execute the conversion processing of the above-described equations (1) and (2), similarly to the conventional basic unit 1.

As shown in FIG. 4, the input unit 1'a-h of the present invention receives a conventional input unit 1'-h which receives an input signal and distributes and outputs the input signal, and receives the distributed output of the input unit 1'-h. A plurality of attribute value calculation units 1′-hq (q
= 1 to n). The attribute value calculation unit 1'-hq newly provided according to the present invention is weighted by a weight value for the basic unit 1-i of the intermediate layer, instead of the input unit 1'-h conventionally provided. Processing will be performed to provide input values.

The newly provided attribute value calculation unit 1'-hq
Performs processing so that the "ambiguity" of the magnitude of the input signal value X input to the input unit 1'-h is digitized according to the membership function of the fuzzy set and output. For example, when the number of units of the attribute value calculation unit 1'-hq connected to the input unit 1'-h is three, the attribute value calculation unit 1'-h1 is, for example, as shown in FIG. When the value of the input signal value X is `` approximately 0.
Membership function indicating that the 3 '(m ₁ in the figure)
Is processed, and a grade value (membership function value) in which the input signal value X is “approximately 0.3” is calculated and output in accordance with the managed function, and the attribute value calculation unit 1′-h2 , A membership function (m ₂ in the figure) indicating that the value of the input signal value X is “approximately 0.5”, and a grade value in which the input signal value X is “approximately 0.5” according to the managed function. treated so as to calculate and outputting, also managing the attribute value calculating unit 1'-h3, membership function representing that the value of the input signal value X is "approximately 0.7" to (m ₃ in the drawing) In addition, the input signal value X is "approximately 0.7" in accordance with the function to be managed, and the grade value is calculated and output, thereby outputting numerical data of the "ambiguity" of the input signal value X. To do To management.

According to the output processing of the attribute value calculation unit 1'-hq, when the input signal value X input to the input unit 1'-h is, for example, "0.3", each of the basic units 1-i of the intermediate layer is processed. , The grade value "1" from the attribute value calculation unit 1'-h1, and the attribute value calculation unit 1'-h2,1 '
The grade value “0” from −h3 is input by being weighted by the learned weight value, and when the input signal value X that is input is, for example, “0.6”,
For each basic unit 1-i of the intermediate layer, a grade value "0" from the attribute value calculation unit 1'-h1 and a grade value "0.33" from the attribute value calculation unit 1'-h2,1'-h3. Is weighted by the learned weight value and input.

Thus, in the present invention, the input unit 1 'of the input layer
-H at least one input unit 1'-h
An attribute value calculation unit 1'-hq is provided, and a grade value which is a semantic expression of the magnitude of the input signal value is calculated according to the provided attribute value calculation unit 1'-hq. In this case, the grade value is distributed to the basic unit 1-i of the intermediate layer instead of the input unit 1'-h.

The configuration for calculating the grade value (membership function value) of the attribute value calculation unit 1'-hq can employ various realization methods. For example, as one method, the sixth
As shown in the figure, there is a method in which a conversion table that describes a membership function is provided, and a truth value of the membership function is calculated by searching the conversion table using an input value as a key. . Here, FIG. 6 (a) shows a configuration for calculating a grade value of the membership function such as "the temperature is low", and FIG. 6 (b) shows a configuration wherein the temperature is normal. FIG. 6C shows a configuration for calculating a grade value of the membership function such as “Temperature is high”.

As another method, an output value y output from the basic unit 1 is used. When (ωx−θ) <0, as shown in FIG. 7 (a), a function shape similar to the membership function described in FIG. 6 (a) is obtained, and (ωx−θ) > 0, paying attention to the fact that the function shape is similar to the membership function described in FIG.
As shown in FIG. 6B, by providing the basic unit 1 and appropriately setting the weight value ω and the threshold value θ, FIG.
An arithmetic operation equivalent to (c) is executed, and a difference y between output values output from the two basic units 1 is calculated. Focusing on the fact that as shown in FIG. 8A, the function shape is similar to the membership function described in FIG. 6B, and as shown in FIG. FIG. 6 shows that by providing the basic units 1 and a subtraction circuit 5 for calculating a difference value between the two output values, by appropriately setting the weight values ω ₁ and ω ₂ and the threshold values θ ₁ and θ ₂ , There is a method that employs a configuration for executing the same arithmetic processing as in b). 7 and 8, even when the hierarchical network structure is composed of hardware components, only the setting of the weight value and the threshold value is performed. The attribute value calculation unit 1'-hq for calculating the grade value of the membership function of the function shape shown in FIG. 8A can be realized, and the function shape shown in FIG. Attribute value calculation unit 1'-hq for calculating the grade value of the membership function of the above.

As a configuration method using hardware components having a hierarchical network structure, there is a method disclosed in Japanese Patent Application No. 63-216865 filed by the present applicant.
No. (filed on Aug. 31, 1988, "Network Configuration Data Processing Apparatus").

That is, the basic unit 1 is, as shown in FIG.
A multiplication type D / A converter 2a for multiplying the output from the basic unit 1 in the preceding layer input via the input switch unit 7 and the weight value held by the weight value holding unit 8, and an integrator for multiplication An analog adder 3a that adds the output value of the type D / A converter 2a and the previous accumulated value to calculate a new accumulated value,
A sample and hold circuit 3b for holding the addition result of the analog adder 3a, a non-linear function generating circuit 4b for non-linearly converting the data held in the sample and hold circuit 3b when the accumulation processing is completed, An output holding unit 5 for holding the analog signal value of the non-linear function generating circuit 4a to be output to the output unit 5, an output switch unit 6 for outputting the final output held by the output holding unit 5 to the analog bus 30, and This is realized by including a control circuit 9 for controlling the processing unit.

The hierarchical network structure is realized by a configuration in which the basic units 1 having this configuration are electrically connected by one common analog bus 30, as shown in FIG. Here, in the figure, 31 is a weight output circuit for giving a weight value to the weight holding unit 8 of the basic unit 1, and 32 is a weight output circuit 31
A weight signal line connecting the output of the hierarchical network to the weight holding circuit 8, an initial signal output circuit 33 for outputting an initial signal serving as an input pattern to an input layer of the hierarchical network, and a synchronization control signal 34 for controlling data transfer. Is transmitted to the weight output circuit 31, the initial signal output circuit 33, and the control circuit 9 of the basic unit 1, and 40 is a main control circuit for transmitting a synchronization control signal. When the configuration shown in FIG. 10 is adopted, a configuration is adopted in which the attribute value calculation unit 1'-hq is provided with another device configuration (not shown).

In the hierarchical network structure of this configuration, the main control circuit 40 sequentially selects the basic units 1 in the preceding layer in a time-series manner, and synchronizes with the selection processing so that the output holding unit 5 of the selected basic unit 1 The final output of the held analog signal is processed to be output to the multiplying D / A converter 2a of the basic unit 1 in the subsequent layer via the analog bus 30 according to a time-division transmission format. Upon receiving this input, the multiplying D / A converter 2a of the subsequent-stage basic unit 1 sequentially selects the corresponding weight value (determined by the back propagation method) and multiplies the input value by the weight value. Processing is performed, and the accumulation processing unit 3 composed of the analog adder 3a and the sample hold circuit 3b sequentially accumulates the multiplied values. Then, the basic unit 1 of the former layer
When all the accumulation processes for are completed, the main control circuit
40 is a non-linear function generating circuit 4a of the basic unit 1 in the subsequent stage.
Is started to calculate the final output, and the output holding unit 5 performs processing so as to hold the final output of the conversion processing result. Subsequently, the main control circuit 40 repeats the same processing for the next subsequent layer using the subsequent layer as a new preceding layer, so that the output pattern corresponding to the input pattern is the output of the hierarchical network structure. It is processed to be output from the layer.

Next, according to the simulation data, by adopting the configuration of the present invention, the learning of the weight value can be performed with higher accuracy.
In addition, it will be described that data can be realized at higher speed, and data processing can be realized with higher accuracy by the learned weight value.

In this simulation, the attribute value calculation unit 1'-
The comparison was made between a model with hq and a model without hq.
The conditions of the simulation are as follows. The configuration of the hierarchical network is such that the input layer is 1 unit, the intermediate layer is 10 units, and the output layer is 1 unit.
-H has three attribute value calculation units 1'-hq (q = 1 to 3) for calculating a grade value in accordance with the membership function shown in FIG. Fig. 11 is a diagram showing the execution of the propagation method (the learning constant ε and the momentum ζ are set to ε = 9.0 and ζ = 0.8, respectively, and the initial values of the weights are set by random number generation means). 9 data consisting of pairs of an input signal value X and a teacher signal value Y shown in FIG.

FIG. 12 shows three attribute value calculation units 1'-hq (q = 3) provided in accordance with the membership function shown in FIG.
1 to 3) show grade values to be output in response to the input of the input signal value X of the teacher signal in FIG.
This figure, for example, when the "0.1111111" of the input signal value X is inputted, the attribute value calculating unit 1'-h1 identified by m ₁ in Figure 5 outputs the "0.1", is identified by m ₂ that outputs the attribute value calculating unit 1'-h2 is "0.0", the attribute value calculating unit 1'-h3 identified by m ₃ indicates that the outputs "0.0".

From this, when the attribute value calculating unit 1'-hq is provided, it is substantially equivalent to a hierarchical network configuration data processing device having three input units 1'-h and using the signal shown in FIG. 13 as a learning signal. It becomes something. On the other hand, when the attribute value calculation unit 1'-hq is not provided,
One input unit 1'-h as it is becomes a hierarchical network configuration data processing device using the signal shown in FIG. 11 as a learning signal.

When the hierarchical network structure of the present invention is used, the simulation data shown in FIG. 14 shows the input signal value X of the learning signal shown in FIG.
Is given, the error value between the output value from the output layer and the corresponding teacher signal value Y (the error value for each of the nine teacher signals and the sum of the difference values thereafter) is determined by the progress of the number of times of learning. This is the data of a simulation performed to investigate how it changes along with. Here, the data in FIG. 14 (a) represents the error value when the number of learnings is 500,
The data in FIG. 4 (b) represents an error value when the number of learnings is 1000 times, and the data in FIG. 14 (c) represents an error value when the number of learnings is 1500 times, and FIG. The data in the parentheses) represent the error value at the time of 1788 times, which is the number of learnings in which the error value has almost converged.

From the simulation data shown in FIG. 14, it can be seen from the simulation data that the total error value was “0.043249” when the number of times of learning was 500, but rapidly decreased with the progress of the number of times of learning by providing the present invention. Thus, it has been confirmed that when the number of times of learning reaches 1788 times, the number decreases to "0.000205". Then, by using the weight value when the number of times of learning is 1788 times, it has been confirmed that the output value from the output unit 1-j can be substantially equal to the same value as the teacher value for all the learning signals. .

FIG. 15 shows simulation data of an error value when a conventional hierarchical network structure to be compared with the data of FIG. 10 is used. In this simulation, the learning signal shown in FIG. 11 is used as described above. Here, the data in FIG. 15 (a) represents an error value when the number of times of learning is 500 times,
The data in FIG. 5 (b) represents an error value when the number of learnings is 1,000, the data in FIG. 15 (c) represents an error value when the number of learnings is 1500, and FIG. ) Represents an error value when the number of learnings is 2,000, and the data in FIG. 15E represents an error value when the number of learnings is 3,000.
The data in FIG. 15 (f) represents an error value when the number of learnings is 4,000, and the data in FIG. 15 (g) represents an error value when the number of learnings is 5,000.

From the simulation data of FIG. 15, if the present invention is not provided, the total error value was “0.031051” when the number of learnings was 500, but gradually decreased with the progress of the number of learnings. Even when the number of times of learning reaches 2,000, the total error value shows a considerably large value of "0.015456", and even when the number of times of learning reaches 5,000, the total error value becomes "0.011816".
It was confirmed that the values did not converge within a realistic range of the number of times of learning.

Thus, from the comparison of the simulation data shown in FIGS. 10 and 15, it can be seen from the comparison of the simulation data shown in FIGS. It was confirmed. As described above, since the number of input units increases formally by including the present invention, the learning time of the weight value should increase in a simple manner. Conversely, the learning time is shortened because the grade value given by the attribute value calculation unit 1'-hq contains information having a larger meaning content. This semantic information enables convergence of learning of the weight value.

FIG. 16 shows the teacher signal shown in FIG. 11 and FIG.
The output value of the simulation data (the X value on the horizontal axis corresponds to the teacher signal) and the output value of the simulation data shown in FIG. Corresponding). From this figure, it can be seen that by using the present invention, when the input signal value of the teacher signal is input to the input layer, an output value that cannot be distinguished from the teacher signal value from the hierarchical network structure can be obtained. Thus, it can be seen that in the conventional method, only an output value with a large error can be obtained despite a long learning time.

Although the illustrated embodiment has been described, the present invention is not limited to this. For example, although the embodiment discloses that the attribute information of the input signal value is calculated using a membership function having a triangular grade value conversion function characteristic,
The conversion function characteristic of the grade value of the membership function may have any shape. The calculation configuration of the attribute information does not need to be limited to the configuration provided according to the network configuration of the integrated configuration, but may be the configuration provided according to the network configuration of another device configuration such as the preprocessing configuration. There may be.

Further, in the embodiment, the attribute value calculation units related to the same group receive input values of the same level from the corresponding input units (that is, the weight value of the internal connection between the attribute value calculation unit and the input unit is “1”). However, since the present invention does not make the absolute value a problem, a weight value other than “1” is assigned to this internal connection, so that the input value and the weight The configuration may be such that the grade value of the membership function is calculated by multiplying the value with the value, or the weight value of the internal connection may be treated as a learning target. .

The present invention is not limited to the one that calculates the attribute information of the input signal value according to the membership function.

Further, the present invention is not limited to a network configuration data processing device having a hierarchical network configuration, and can be applied to a network configuration data processing device having a network configuration other than the hierarchical network configuration.

〔The invention's effect〕

As described above, according to the present invention, in a network configuration data processing device, even when learning of a weight value of a network structure does not converge in a conventional method, learning of a weight value can be quickly realized. become. Thus, the data processing function of the network configuration data processing device can be made more accurate. In addition, since the learning speed can be increased, the network configuration data processing device can be made more practical.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram of the principle of the present invention, FIG. 2 is an explanatory diagram of an attribute function, FIG. 3 is an embodiment of a hierarchical network according to the present invention, FIG. One embodiment of such an input unit, FIG. 5 is an example of a membership function assigned to an attribute value calculation unit, and FIGS. 6, 7, and 8 are attribute values for calculating a grade value of the membership function. FIG. 9 is an explanatory view of an example of a configuration adopted by a calculation unit, FIG. 9 is an embodiment of a basic unit, FIG. 10 is an embodiment of a hierarchical network, FIG. 11 is an explanatory view of a teacher signal used for simulation, and FIG. FIG. 13 is an explanatory diagram of a grade value output from the attribute value calculating unit, FIG. 13 is an explanatory diagram of a teacher signal used for simulation, FIG. 14 and FIG. 15 are explanatory diagrams of simulation data, and FIG. Data conversion processing Illustration of comparison, FIG. 17 is a basic block diagram of the basic unit, FIG. 18 is a basic configuration diagram of a hierarchical network. In the figure, 1 is a basic unit, 1 'is an input unit, 2 is a multiplication processing unit, 3 is an accumulation processing unit, 4 is a threshold processing unit, 10 is a network configuration data processing device, 11 is a network structure unit, and 12 is an internal unit. This is a state value storage unit.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Asahi Kawamura 1015 Uedanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture Inside Fujitsu Co., Ltd. 72) Inventor Kazuo Asakawa 1015 Uedanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture Within Fujitsu Limited (72) Inventor Shigenori Matsuoka 1 Fujimachi, Hino-shi, Tokyo Fuji Facom Control Co., Ltd. (72) Inventor Hiroyuki Okada Tokyo 1 Fujimachi, Hino-shi, Fuji Faccom Control Co., Ltd.

Claims (2)

(57) [Claims] 1. A method for receiving one or more inputs and an internal state value to be multiplied with respect to the input to obtain a product sum value, and obtaining an output value by subjecting the product sum value to a function conversion. In addition to being constituted by the internal connection of the basic units of
A network structure having one or more input units for receiving an input signal value of the input pattern and distributing the input signal value to a corresponding basic unit, wherein the network structure is defined by a network structure of the network structure and an internal state value of the network structure; In a network configuration data processing apparatus for calculating and outputting an output pattern corresponding to an input pattern according to a data conversion function to be performed, at least one of the input units divides a space of a signal value input to the own unit. Assigning an attribute function that defines an attribute value of an input signal value to each of the divided spaces, and, when receiving an input signal value, calculating attribute values for the number of the attribute functions according to the attribute function. Characterized in that it is configured to input values to the corresponding basic units. Configuration data processing apparatus. 2. An input unit for distributing an input signal, receiving one or more inputs from a preceding layer and an internal state value to be multiplied to the input to obtain a sum of products value.
A basic unit that obtains an output value by converting the product-sum value into a function, and a plurality of input units serving as an input layer, and a plurality of the basic units serving as an intermediate layer serving as one or more intermediate units. And one or a plurality of the above basic units as an output layer, and an internal portion between the input layer and the first intermediate layer, between the intermediate layers, and between the last intermediate layer and the output layer. In a network configuration data processing apparatus comprising a network structure configured to form a connection and to set an internal state value corresponding to the internal connection, at least one of the input units is input to its own unit. Divides the space of the incoming signal values, assigns an attribute function that defines the attribute value of the input signal value to each of the divided spaces, and, when receiving the input signal value, according to the attribute function. A network configuration data processing apparatus characterized in that it is configured to calculate attribute values for the number of attribute functions and to input the calculated values to the basic unit of the intermediate layer at the forefront stage.