JPH06168219A - Fuzzy structure type neuro computer and financial consultant information collecting method - Google Patents

Abstract

PURPOSE:To provide the fuzzy structure type neuro computer for enabling exact consulting to data processing and the financial consultant information collecting method for exactly collecting financial consultant information. CONSTITUTION:A fuzzy structure type neuro computer 100 provided with network structure according to the fuzzy rule of a data processing object calculates the differentiated value of an antecedent part membership function from the function value of the antecedent part membership function in the fuzzy rule outputted from the network structure and an internal state value for providing the antecedent part membership function and outputs the calculated value as consultant information. On the other hand, the internal state value of the fuzzy structure type neuro computer 100, to which the fuzzy rule of bond ranking is mapped, is learnt by using the plural kinds of bond ranking information, and the financial consultant information outputted from these respective internal state values is collected.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fuzzy structural neurocomputer that executes data processing in accordance with a data conversion function of a network structure structured by fuzzy rules, and a bond rating using this fuzzy structural neurocomputer. Financial consulting information collection method that collects financial consulting information necessary for financial consulting, especially for fuzzy structured neurocomputers that enable accurate consulting for data processing and accurate financial consulting The present invention relates to a financial consult information collecting method that enables collection of financial consult information that can realize

The applicant of the present invention has disclosed a fuzzy structure type neuro computer in Japanese Patent Application No. 2-60256 (Title of the invention: Hierarchical network configuration data processing device and data processing system). The fuzzy structure-type neurocomputer has a configuration of executing data processing according to a data conversion function of a network structure structured by fuzzy rules.

According to this configuration, the fuzzy structural neurocomputer has a configuration in which the weakness of the neurocomputer is complemented by the fuzzy model and the weakness of the fuzzy model is complemented by the neurocomputer by fusing the fuzzy model and the neurocomputer. It is characterized by being taken. That is, the explanation function that is lacking in the neurocomputer, such as what kind of data conversion processing is being performed, is complemented by a fuzzy model that is easy to model, and at the same time, it is difficult to adjust the membership function etc. It is characterized by complementing the weaknesses it has with a neurocomputer that is easy to learn.

In order to improve the practicality of this fuzzy structure type neuro computer, it is necessary to provide a structure that enables accurate consulting for data processing.

The applicant of the present invention is Japanese Patent Application No. 4-137380 (Title of Invention: Bond Rating Determining Device and Financial Consulting Method) and a bond rating determining device using this fuzzy structural neurocomputer, and the bond thereof. A financial consulting method for executing financial consulting for bond ratings using a rating determination device is disclosed.

[0006] In order to improve the practicality of financial consulting for such bond rating, it is necessary to construct a structure capable of collecting financial consult information that can realize appropriate financial consulting. .

[0007]

2. Description of the Related Art Next, a fuzzy structure type neurocomputer disclosed by the present applicant in Japanese Patent Application No. 2-60256 will be described.

As shown in FIG. 14, this fuzzy structure type neurocomputer is connected to an input layer 10 composed of linear neurons 1 for the input variables and the input layer 10 by an internal connection to which internal state values are assigned. A first stage layer 11 composed of a plurality of combinations of sigmoid neurons 2 and linear neurons 1;
Combined with 1 and the internal join where the internal state value is assigned,
A second-stage layer 12 composed of a plurality of sigmoid neurons 2 and a third-stage layer 13 composed of a plurality of linear neurons 1 connected to the second-stage layer 12 by an internal connection to which internal state values are assigned. And a fourth stage layer 14 composed of a plurality of linear neurons 1 connected to the third stage layer 13 by an internal connection to which an internal state value is assigned, and a fourth stage layer 14 and an internal state value assignment. The output signal of the fifth-stage layer 15 is prepared separately from the fifth-stage layer 15 composed of two linear neurons 1 and the hierarchical network structure described above. And a conclusion value determining means 16 for receiving.

The fuzzy structure-type neurocomputer shown in FIG. 14 has a hierarchical network structured so as to map the fuzzy rules shown in FIG. Further, the linear neuron 1 outputs an output signal value y represented by y = Σω _i x _i, where x is an input signal value and ω is an internal state value assigned to the internal coupling of the input signal value x. The sigmoid neuron 2 inputs the input signal value x,
The internal state value to be assigned to the internal coupling of the input signal value x omega, if the threshold of the sigmoid function θ, y = 1 / (1 + exp (-Σω i x i + θ)) output signal value y represented by It operates so as to output.

The first layer 11 of the fuzzy structure type neurocomputer, when an input signal (X, Y in the example of FIG. 14) is given, is a function of the antecedent part membership function of the fuzzy rule of the input signal. Operates to calculate a value.

The calculation process of the function value of the antecedent part membership function is realized by the following configuration. That is, the antecedent part membership function is defined as having a function shape as shown in FIG. 16, while the sigmoid neuron 2 has input / output signal characteristics as shown in FIG. 17 by the signs of ω and θ. Have

From this, the antecedent membership function such as "X is small" in FIG. 16 has a negative internal state in the internal connection between the linear neuron 1 of the input layer 10 and the sigmoid neuron 2 of its own layer. This can be realized by assigning the value ω, the sigmoid neuron 2 calculates a negative threshold value θ, and assigning the internal state value “1” to the internal connection with the subsequent linear neuron 1. In addition, the antecedent part membership function such as “X is big” in FIG. 16 assigns a positive internal state value ω, calculates a positive threshold value θ, and outputs a value between the subsequent linear neuron 1. This can be realized by assigning the internal state value "1" to the internal coupling of.

Further, the antecedent part membership function such as "X is middle" in FIG. 16 allocates a negative internal state value ω and also calculates a negative threshold value θ and a sigmoid neuron 2 and a positive internal value. A sigmoid neuron 2 for allocating a state value ω and calculating a positive threshold value θ is prepared, and these two sigmoid neurons 2 are prepared.
In order to calculate the difference value of the output values of the internal state values “1” and “−
It can be realized by assigning 1 ”.

The second-stage layer 12 of the fuzzy structure-type neurocomputer is composed of sigmoid neurons 2 provided in association with each fuzzy rule, and a function value for the antecedent part proposition is given from the first-stage layer 11. Sometimes it works to calculate the grade value for a fuzzy rule.

The calculation process of the grade value for this fuzzy rule can be realized by various network connections. For example, it is realized by the following configuration. That is, the antecedent part proposition of the fuzzy rule is "X is small
And the condition "(X is small
) And (Y is small) ".
For the former propositional description, for example, an internal state value of “4.98” is assigned to the internal connection between the linear neuron 1 of the first stage layer 11 and the sigmoid neuron 2 of its own layer, which are in a propositional description relation. At the same time, it can be realized by the sigmoid neuron 2 of its own layer calculating the threshold value of "-2.49". By assigning this internal state value / threshold value, the second
The sigmoid neuron 2 of the stair layer 12 outputs the function value output from the linear neuron 1 of the first stair layer 11 as it is to the third
The grade value for the fuzzy rule is calculated by outputting the grade value to the stage layer 13.

On the other hand, with respect to the latter propositional description, for example, the internal connection between the two linear neurons 1 of the first stage layer 11 and the sigmoid neuron 2 of its own layer, which are in the propositional description relation, is called “7.0”. It can be realized by allocating the internal state value and the sigmoid neuron 2 of its own layer calculates a threshold value of "10.5". This internal state value /
By assigning the threshold value, the sigmoid neuron 2 of the second layer 12 has a function value μ _A output by one linear neuron 1 of the first layer 11 and a function value μ _B output by the other linear neuron 1. The marginal product is “0 ∨ (μ _A + μ _B −
1) ”is roughly calculated and output to the third stage layer 13 to calculate the grade value for the fuzzy rule.

The third-stage layer 13 of the fuzzy structure-type neurocomputer is composed of a linear neuron 1 provided in association with the consequent part of the fuzzy rule,
When the grade value for the fuzzy rule is given from the stage layer 12, it operates so as to calculate the proposition value for the consequent part proposition of the fuzzy rule.

The process of calculating the proposition value for the consequent part proposition is based on the fuzzy rule for the internal connection between the sigmoid neuron 2 of the second stage layer 12 and the linear neuron 1 of the self layer in the propositional description relation. It is realized by assigning an internal state value indicating the degree of importance (weight value).

The fourth stage layer 14 of the fuzzy structure type neurocomputer is composed of an appropriate number of linear neurons 1 provided in units of the same output variable of the fuzzy rule, and the proposition value is given from the third stage layer 13. At the same time, it operates so as to reduce the consequent part membership function of the fuzzy rule according to the proposition value, and to calculate the functional sum of the consequent part membership functions for the same reduced output variables.

The calculation processing of the reduction function sum for the consequent part membership function is realized by the following configuration. That is, as shown in FIG. 18, the consequent part membership function is finely divided according to the number of the linear neurons 1 of the fourth layer 14 to specify the function value value α _i of each section, and the specified function value It is realized by allocating α _i as the internal state value of the internal connection between the linear neuron 1 of the third layer 13 and the linear neuron 1 of the own layer.

For example, for the consequent part proposition "Z is small", the function value for interpolating the consequent part membership function of "Z is small" is transferred to the linear neuron 1 of the third layer 13. Assigned as the internal state value of the inner join of
For the consequent part proposition "Z is middle", "Z i
The function value that interpolates the consequent part membership function of "s middle" is assigned as the internal state value of the internal connection with the linear neuron 1 of the third stage layer 13, and the consequent part proposition "Z is big" is assigned. By assigning the function value for interpolating the consequent part membership function of “Z is big” as the internal state value of the internal connection with the linear neuron 1 of the third layer 13, Each consequent part membership function of the fuzzy rule is contracted according to the proposition value given by, and the functional sum of consequent part membership functions for the contracted same output variables Z is obtained.

The fifth stage layer 15 of the fuzzy structure type neurocomputer is composed of two linear neurons 1 for the same output variable, and the consequent membership function reduced from the linear neuron 1 of the fourth stage layer 14 When the functional sum of the above is given, it operates so as to calculate two centroid derived values used for calculating the centroid value of the functional sum.

The calculation processing of the derived value of the center of gravity is realized by the following configuration. That is, the internal connection between one linear neuron 1 of the fifth layer 15 and the linear neuron 1 of the fourth layer 14 is assigned an internal state value that increases from 0 to 1 at equal intervals. , The inner connection between the other linear neuron 1 and the linear neuron 1 of the fourth stage layer 14 can be realized by assigning internal state values that increase from -1 to 0 at equal intervals.

When the internal state values are assigned in this way, for example, when the number of linear neurons 1 in the fourth stage layer 14 is 6, the six linear neurons 1 are C _i.
For example, one of the linear neurons 1 of the fifth stage layer 15 outputs Z ₁ = 0 · C ₁ +0.2 · C ₂ +0.4 · C ₃ +0.6 · C ₄
+0.8 · C ₅ +1 · C ₆ other linear neuron _{1, Z 2 = -1 · C 1} -0.8 · C 2 -0.6 · C 3 -0.4 · C
That _{4 -0.2 · C 5 -0 · C} 6 calculates the centroid derived value.

When Z = Z ₁ / (Z ₁ -Z ₂ ) is calculated by using these two derived values of the center of gravity Z ₁ and Z ₂ , Z = Z ₁ / (C ₁ + C ₂ + C ₃ + C ₄ + C ₅ + C ₆ ) will be calculated, which represents the centroid value of the sum of functions represented by grade (z), such as Z = ∫grade (z) zdz / ∫grade (z) dz. There is. From now on, the linear neuron 1 of the fifth layer 15
The calculated value calculated by is the center-of-gravity derived value used for calculating the center-of-gravity value of the functional sum of the reduced consequent part membership functions output from the fourth layer 14.

The conclusion value deciding means 16 provided in the fuzzy structure type neurocomputer operates so as to calculate the conclusion value which becomes the final output of the fuzzy rule when the two barycentric derivation values are given from the fifth stage layer 15. .

As described above, the calculation process of the conclusion value is performed by the two center-of-gravity derivation values Z _1, Z ₂ output from the fifth layer 15.
It is realized by calculating Z = Z ₁ / (Z ₁ −Z ₂ ) using. The conclusion value determining means 16 can also be configured by a network structure.

As described above, the fuzzy structure-type neurocomputer disclosed by the present applicant in Japanese Patent Application No. 2-60256 is characterized by having a hierarchical network structure structured according to fuzzy rules. When the hierarchical network structure is initialized according to the fuzzy rule of the data processing target, and subsequently the teacher signal related to the input / output signal of the data processing target is obtained, the learning processing device that implements the backpropagation method, etc. So that the output signal that is output when the input presentation signal is input becomes the output presentation signal of the teacher signal, the internal state value / threshold of the hierarchical network structure to be adjusted (the input layer 10 and the first stage layer 11). , The threshold value of the sigmoid neuron 2 of the first stage layer 11, the internal state value between the second stage layer 12 and the third stage layer 13, By learning the internal state value) between the layer 13 and the fourth-stage layer 14 is intended to adopt a configuration that build a data processing of the data processed as performing data processing apparatus.

In the fuzzy structure-type neurocomputer configured as described above, conventionally, the weight value of the fuzzy rule obtained by learning (the internal state value between the second layer 12 and the third layer 13) By evaluating the size of
Whether consulting for data processing is performed, or whether consulting for data processing is performed by evaluating the magnitude of the grade value (output value of the second layer 12) for the fuzzy rule. A method has been adopted in which consulting for data processing is performed by evaluating the size of the multiplication value of the weight value and the grade value.

For example, in the case of constructing a fuzzy structural neurocomputer as a bond rating determination device, when consulting on how to improve financial data in order to raise the bond rating of a certain company, Among the fuzzy rules that describe the consequent part proposition to increase the bond rating, specify the one with a large weight value, the one that outputs a large grade value, the weight value and the grade value. The method of consulting was adopted by identifying the one with a large multiplication value of and operating the financial data described in the antecedent part proposition of the identified fuzzy rule.

[0031]

However, according to such a conventional technique, although the degree of influence of the data conversion function on the current input value (value of the input variable) indicated by the data processing target is known, Since the rate of change of the output value with respect to the input value is not taken into consideration, there is a problem that the input variable to be operated to obtain the desired output value cannot be accurately grasped. That is, according to such a conventional technique, there is a problem that accurate consulting for data processing cannot be executed.

And, Japanese Patent Application No. 4-137 disclosed by the present applicant.
The invention of No. 380 (name of the invention: bond rating determination device and financial consulting method) has the drawback that sufficient financial consulting cannot be executed due to this problem, and a structure that allows consideration of differences in bond rating agencies. Since it did not disclose it, there was a drawback in that it could not carry out sufficient financial consulting.

The present invention has been made in view of the above circumstances, and is a fuzzy structure-type neurocomputer that executes data processing according to a data conversion function of a network structure structured by fuzzy rules. With the aim of providing a new fuzzy structural neurocomputer that enables accurate consulting, all financial consulting information necessary for financial consulting on bond ratings using this fuzzy structural neurocomputer. The purpose of the present invention is to provide a new financial consult information collection method that enables the collection of financial consult information that enables appropriate financial consulting when collecting information.

[0034]

FIG. 1 shows a principle configuration of a fuzzy structure type neurocomputer of the present invention, and FIG. 2 shows a principle configuration for realizing a financial consult information collecting method of the present invention. To do.

In FIG. 1, reference numeral 100 is a fuzzy structure-type neurocomputer constructed according to the present invention. The fuzzy structure type neuro computer 100 includes a network configuration data processing means 20, a conclusion value determining means 16,
The internal state value management means 21, the calculation means 22, the multiplication means 23, and the output means 24 are provided. Here, as described in FIG. 14, 1 is a linear neuron and 2 is a sigmoid neuron.

The network configuration data processing means 20 is
The input layer 10, the first stage layer 11, the second stage layer 12, the third stage layer 13, the fourth stage layer 14 and the fifth stage layer 15 described in FIG. Perform data processing according to the data conversion function of the structured network structure.

As described with reference to FIG. 14, the conclusion value determining means 16 is composed of the network structure or the non-network structure, and the final output of the fuzzy rule is the final output of the fuzzy rule from the centroid derived value calculated by the network configuration data processing means 20. Determine the value (output value). Internal state value management means 21
Manages the internal state value assigned to the internal connection of the network structure portion of the network configuration data processing means 20.

The calculation means 22 uses the data indicated by the network configuration data processing means 20 to execute the prescribed calculation processing. The multiplication unit 23 executes a calculation process of a multiplication value of the calculation value calculated by the calculation unit 22 and the data indicated by the network configuration data processing unit 20. Output means 24
Outputs the calculated value calculated by the calculating unit 22 and the multiplied value calculated by the multiplying unit 23 as consult information to the outside.

Reference numeral 200 denotes a teacher signal management device which manages the input / output data relationship of the data processing target as a teacher signal. Reference numeral 300 denotes a learning processing device, and when the input value of the teacher signal managed by the teacher signal management device 200 is input to the network configuration data processing means 20, the output value of the paired teacher signal from the conclusion value determination means 16 In order to be output, the internal state value of the learning target of the network configuration data processing means 20 is learned according to an algorithm such as the back propagation method, and is stored in the internal state value management means 21. It is a thing.

In FIG. 2, the same components as those described in FIG. 1 are designated by the same symbols. 200a-i (i = 1
~ N) is a teacher signal management device provided for, for example, a bond rating institution, and corresponds to the financial data of a number of companies and the bond rating information for these companies rated by each bond rating institution. The data is managed as a teacher signal. In addition to this, the teacher signal management device 20
0a-i may be provided for each year, for example, and may manage correspondence data between financial data of a large number of companies for each year and bond rating information for these companies as a teacher signal.

The learning processing device 300a sequentially selects the teacher signal management devices 200a-i, and outputs the financial data of the teacher signals managed by the selected teacher signal management device 200a-i to the network configuration data processing means 20. In order to ensure that the bond value information of the paired teacher signal is output from the conclusion value determining means 16 when inputting into the, the learning target of the network configuration data processing means 20 according to an algorithm such as the backpropagation method. The internal state value is learned.

Reference numeral 400 denotes a financial consult information collecting device, which collects financial consult information necessary for financial consulting for bond rating. Here, in the configuration of FIG. 2, a plurality of fuzzy structural neurocomputers 100 may be provided, for example, for a bond rating agency.

[0043]

In the present invention, when the user adopts the configuration of executing the data processing of the data processing target according to the fuzzy structure type neuro computer 1, first, the fuzzy rule for the data processing target is generated.

Next, the user uses the fuzzy rule generated in this way as the network configuration data processing means 20.
To the network structure part of. That is, in order to realize the calculation of the function value of the antecedent membership function of the fuzzy rule, the internal state value assigned to the internal connection to the sigmoid neuron 2 of the first stage layer 11 is set and the sigmoid The threshold value that the neuron 2 calculates is set, and this sigmoid neuron 2
The internal state value of the internal connection between the following linear neuron 1 is set to "1" or "-1" in conformity with the function shape of the antecedent part membership function.

Then, in accordance with the antecedent part of the fuzzy rule, the internal connection with respect to the internal connection between the sigmoid neuron 2 of the second stage layer 12 and the linear neuron 1 of the first stage layer 11 in the propositional description relation The state value is set, and the threshold value calculated by the sigmoid neuron 2 is set. As described above, for example, for the antecedent part proposition without "and", the internal state value is set to "4.98" and the threshold value is set to "-2.49".

Then, the internal state value indicating the degree of importance (weight value) of the fuzzy rule is set to the second value in the propositional description relation.
An internal connection is set between the sigmoid neuron 2 of the layer 12 and the linear neuron 1 of the third layer 13. That is, when the importance of rule 1 is “1.0”, the sigmoid neuron 2 of the second stage layer 12 to which rule 1 is assigned and the sigmoid neuron 2 of the second part layer 13 to which the consequent part of the rule 1 is assigned are The internal connection with the linear neuron 1 is set to "1.0".

Then, the partition function value of the consequent part membership function of the fuzzy rule is set as the internal state value, and the third
The internal connection between the step layer 13 and the fourth step layer 14 is set.
In this way, the fuzzy rule regarding the data processing target is mapped to the network structure portion of the network configuration data processing means 20.

Subsequently, the user activates the learning processing device 300. When activated in this way, the learning processing device 300 forms a pair from the conclusion value determination means 16 when inputting the input value of the teacher signal managed by the teacher signal management device 200 to the network configuration data processing means 20. In order to output the output value of the teacher signal, learning of the internal state value to be learned is executed according to the algorithm of the back propagation method. That is, the first stage layer 11
Internal state values assigned to the internal connection to the sigmoid neuron 2 and the threshold values calculated by these sigmoid neurons 2, and the internal state values assigned to the internal connection to the linear neuron 1 of the third layer 13 And the internal state value assigned to the internal connection of the fourth-stage layer 14 to the linear neuron 1.

By this learning process, the fuzzy structure-type neurocomputer 100 composed of the network configuration data processing means 20 and the conclusion value determining means 16 receives an unlearned input value for data processing,
It operates so as to output the output value associated with the input value.

When the learning process of the learning processing device 300 is completed, the user inputs the input value of the data processing target to be consulted to the fuzzy structural neurocomputer 1.
Enter 00.

At this time, the calculation means 22 uses the first layer 11
The function value of the antecedent part membership function of the fuzzy rule and the corresponding internal state value assigned to the internal connection to the sigmoid neuron 2 of the first layer 11 (the internal part that realizes the antecedent part membership function). (Which is a state value), the differential value of the antecedent part membership function is calculated, and in response to this calculation result, the multiplication means 23 uses the differential value calculated by the calculation means 22 and the derivation source of this differential value. A multiplication value with a weight value indicated by a certain fuzzy rule (internal state value between the second stage layer 12 and the third stage layer 13) is calculated.

Alternatively, the calculation means 22 may be the first layer 11
Function value of the antecedent part membership function of the fuzzy rule and the desired value of the function value (for example, “1” when you want to obtain an output value that makes the consequent part membership function described by the fuzzy rule large. , To obtain a smaller output value, a difference value with respect to a value such as “0” is calculated, and in response to this calculation result, the multiplication means 23 calculates the difference value calculated by the calculation means 22 and the difference value. The multiplication value with the weight value indicated by the fuzzy rule that is the derivation source of is calculated.

Here, the calculating means 22 calculates the function value of the antecedent membership function of the fuzzy rule output by the first-stage layer 11 and the grade value for the fuzzy rule output by the second-stage layer 12. Since and show a strong correlation,
In place of the function value of the antecedent membership function of the fuzzy rule output from the first layer 11, the grade value for the fuzzy rule output from the second layer 12 may be used.

The output means 24 outputs the calculated value calculated by the calculating means 22 or the multiplied value calculated by the multiplying means 23 to the outside as consult information. The consult information output in this manner is output by the fuzzy structural neurocomputer 100 and the input value input to the fuzzy structural neurocomputer 100 when the calculating unit 22 adopts the above-described configuration for calculating the differential value. The calculation means 22 represents the rate of change from the output value.
When adopting the configuration for calculating the difference value described above, the user indicates that the output value output from the fuzzy structure type neurocomputer 100 and the deviation of the output value from the desired output are expressed. According to the consult information, it becomes possible to more accurately grasp the input variable to be manipulated to obtain the desired output and the manipulated variable thereof than in the past.

As described above, according to the present invention, it is possible to realize the fuzzy structure type neurocomputer 100 which enables accurate consulting for data processing.

In the present invention, when the user collects financial consult information for bond ratings, first of all,
Generate fuzzy rules that describe the correspondence between financial data and bond ratings.

Next, the user applies the fuzzy rules thus generated to the network configuration data processing means 20.
To the network structure part of. This mapping process
It will be executed as described above.

Subsequently, the user selects the learning processing device 300a.
To start. When activated in this way, the learning processing device 300a sequentially selects the teacher signal management devices 200a-i and sets the financial data of the teacher signals managed by the selected teacher signal management device 200a-i as network configuration data. According to the algorithm of the backpropagation method, the bond value information of the teacher signal to be paired is output from the conclusion value determination means 16 when inputting to the processing means 20. Learning is executed and the learned internal state value is notified to the financial consult information collecting device 400. That is, the internal state values assigned to the internal connections of the first-stage layer 11 to the sigmoid neuron 2 and the threshold values calculated by these sigmoid neurons 2 are learned, and the internal state of the third-stage layer 13 to the linear neuron 1 is learned. Since the internal state values assigned to the connection and the internal state values assigned to the internal connection to the linear neuron 1 of the fourth layer 14 are learned and the internal state values are notified to the financial consult information collecting device 400. is there.

By this learning process, the fuzzy structural neurocomputer 100, which is composed of the network configuration data processing means 20 and the conclusion value determining means 16, receives the financial data for which the bond rating has not been executed. , The bond rating information indicated by the financial data will be output.

When the learning process of the learning processing device 300a is completed, the user activates the financial consult information collecting device 400. When activated in this way, the financial consultation information collecting apparatus 400 sequentially selects the internal state values of the respective teacher signal management apparatuses 200a-i notified from the learning processing apparatus 300a, and selects the selected internal state values. While setting in the internal state value management means 21, the financial data of the financial consulting target is input to the fuzzy structural neuro computer 100 in synchronization with the setting.

In response to this input, as described above, the output means 24 outputs the calculated value calculated by the calculating means 22 or the multiplied value calculated by the multiplying means 23 to the outside as consult information. Consult information collection device 400
Collects the output consult information as financial consult information. The financial consult information collected in this way represents financial consult information regarding bond ratings indicated by different bond rating agencies and financial consult information regarding bond ratings indicated by different years. It becomes possible to carry out appropriate financial consulting by referring to. Moreover, the financial consult information thus collected has a fuzzy structural neurocomputer 1 when the calculating means 22 adopts the above-mentioned differential value calculating structure.
00 represents the rate of change between the financial data input to 00 and the bond rating information output by the fuzzy structural neurocomputer 100, and when the calculating unit 22 has a configuration for calculating the above-mentioned difference value, Since it represents the magnitude of the deviation between the bond rating information output by the fuzzy structural neurocomputer 100 and the desired output of the bond rating information, the user can more accurately than before according to the financial consult information. It becomes possible to grasp the financial data type to be operated to obtain the desired output and the data operation amount.

When a plurality of fuzzy structure-type neurocomputers 100 are provided for the teacher signal management devices 200a-i, each fuzzy structure-type neurocomputer 10 is subjected to the learning process of the learning processing device 300a.
Since 0 allows the learned internal state value to be set in the internal state value management means 21 as it is, the financial consult information collecting device 400 does not perform the internal state value setting process, By directly inputting financial data to the structural neuro computer 100, financial consult information can be collected.

As described above, according to the present invention, it becomes possible to collect financial consult information capable of realizing accurate financial consulting for bond rating.

[0064]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The applicant of the present invention is the invention of Japanese Patent Application No. 4-137380 (the title of the invention: a bond rating determination device and a financial consulting method), and a fuzzy structural neurocomputer 1
It has been disclosed that the bond rating determination device is constructed using 00.

First, the fuzzy structure type neurocomputer 100 of the present invention will be described in detail in accordance with the embodiment applied to the bond rating determining apparatus. In the invention of Japanese Patent Application No. 4-137380, by examining the correspondence between the financial data of 56 companies and the bond rating, a fuzzy rule that describes the correspondence between the financial data and the bond rating is as follows. The higher the profit, the higher the bond rating. (2) If ordinary income is normal, the bond rating is normal. (3) If the ordinary income is small, the bond rating is low. The most basic three basic classification rules are: (1) If the equity capital is large, the bond rating is high. (2) If the equity capital is small, the bond rating is low. (3) If the interest coverage ratio is large, the bond rating is high. (4) If the interest coverage ratio is small, the bond rating is low. (5) If the long-term debt composition ratio is low, the bond rating is high. (6) If the long-term debt composition ratio is high, the bond rating is low. (7) If the equity ratio is low, the bond rating is low. We obtained 7 modified classification rules that complement the basic classification rules.

FIG. 3 shows a list of these fuzzy rules. Here, in the figure, “p” indicates that the antecedent part proposition is “high / large”, “q” indicates that the antecedent part proposition is “normal”, and “r” is the antecedent. It means that the divisional proposition is "low / small". Also, as shown in the figure, "1.0" is set as the importance of the rule for the basic classification rule, and for the modified classification rule,
We set "0.2" as the importance of the rule.

When the fuzzy structure-type neurocomputer 100 is constructed as a bond rating determining device, the network configuration data processing means 20 follows the fuzzy rule to construct a network structure (of the fourth-tier layer 14) as shown in FIG. The number of linear neurons 1 is 10). Note that p, q, r in the figure
Respectively correspond to the propositions p, q, and r in FIG.

In the invention of Japanese Patent Application No. 4-137380, the membership functions of these fuzzy rules are shown in FIG.
It was defined by what is shown in. Here, "decr" in the figure is shown in FIG.
6A represents a decreasing function as shown in FIG. 6A, and “tri” represents a triangular function as shown in FIG.
"cr" represents an increasing function as shown in FIG.
Further, “a” in the figure represents the coordinate position of the turning point a of the membership function shown in FIG. 6, and “b” in the figure.
Represents the coordinate position of the turning point b shown in FIG.
"C" in the figure represents the coordinate position of the turning point c shown in FIG.

When the user constructs the bond rating determination device using the fuzzy structure-type neurocomputer 100, first, the antecedent part membership function of the generated fuzzy rule is transferred to the network of the network configuration data processing means 20. Map to the first layer 11 of the structure.

Regarding this mapping processing, Japanese Patent Application No. 4-13738
It is executed by the method disclosed in the specification of No. 0, and is assigned to the internal connection to the sigmoid neuron 2 of the first layer 11 of the network configuration data processing means 20 according to this mapping processing in FIG. The different internal state values and the threshold values to be calculated by the sigmoid neuron 2 are shown. Here, in the figure, "RK1, R
K2, RK3, RJ1, RJ2, RI1, RI2, RC
1, RC2, RR1 ”represents the sigmoid neuron 2 provided in association with each rule shown in FIG.

On the other hand, as described with reference to FIG. 1, the consequent part membership function of the generated fuzzy rule is the same as that of the third layer 13 and the fourth layer 14 of the network structure of the network configuration data processing means 20. It will be mapped as the set value of the internal state value of the internal connection between.

Regarding this mapping process, Japanese Patent Application No. 4-13738
As disclosed in the specification of No. 0, the bond rating values are “0.1”, “0.2”, “0.3”, “corresponding to the fact that the number of the linear neurons 1 in the fourth stage layer 14 is 10. 0.4 "," 0.5 ",
Specify the function value of each consequent membership function when "0.6", "0.7", "0.8", "0.9", "1.0",
This is executed by setting the specified function value as the internal state value of the internal coupling between the third layer 13 and the fourth layer 14.

In this way, the fuzzy rule describing the correspondence between the generated financial data and the bond rating is mapped onto the network structure of the network configuration data processing means 20.

When the user maps the fuzzy rules on the network structure of the network configuration data processing means 20, the user can apply the fuzzy rules to the teacher signal management device 200 as described in the invention of Japanese Patent Application No. 4-137380. After registering the paired data of the financial data of the five items of 56 companies used for the generation and the rating information by the bond rating agency as a teacher signal, the learning processing device 300 is started. At this time,
The financial data is converted and registered according to the conversion format disclosed in the specification of Japanese Patent Application No. 4-137380, and the rating information is
It will be converted into a bond rating value as shown in FIG. 8 and registered.

When the learning processing device 300 is started, when the financial data of the teacher signal managed by the teacher signal management device 200 is input to the fuzzy structure type neurocomputer 100, the conclusion value determining means 16 makes a pair of bonds. In order to output the rating value, the internal state values assigned to the internal connections to the sigmoid neurons 2 of the first stage layer 11 and the sigmoid neurons 2 are calculated according to the backpropagation algorithm. By learning the threshold value, the function shape of the antecedent part membership function is tuned, and the internal state value assigned to the internal connection to the linear neuron 1 of the third stage layer 13 is learned, The importance (weight value) of each fuzzy rule is tuned, and the internal connection to the linear neuron 1 of the fourth layer 14 is performed. Ri internal state value to be attached it to continue to learn, continue to tune the function shape of the consequent membership functions.

As described above, the learning processing device 300 automatically tunes the fuzzy rule mapped to the network configuration data processing means 20 to a more accurate one.

By this learning processing, the bond rating determination device of the fuzzy structural neurocomputer 100, which is composed of the network configuration data processing means 20 and the conclusion value determination means 16, follows the neurocomputer operation.
When the financial data for which the bond rating is not executed is input, the bond rating information indicated by the financial data is output.

9 and 10 show the internal state value and the threshold value of the learning result obtained when the learning is performed 1300 times. Here, FIG. 9 is an internal state value defining the consequent part membership function, the broken line represents the internal state value after learning execution, and the solid line represents the internal state value before learning execution. It is a thing.

When the learning of the learning processing device 300 is completed, the user inputs the financial data of the company, which he / she wishes to consult the bond rating, to the bond rating determining device of the fuzzy structural neurocomputer 100. That is,
In order to improve the rating of a bond, financial data of a company that wishes to consult which financial data item should be improved and how much should be input is input to the bond rating determination device of the fuzzy structural neurocomputer 100.

In this way, when the financial data to be consulted is input, the calculation means 22 makes the first
Using the function value M of the antecedent part membership function output from the linear neuron 1 of the stage layer 11 and the internal state value ω that realizes the antecedent part membership function, “M (1-M) ω”
The differential value of the antecedent membership function of each fuzzy rule is calculated by calculating the value of the arithmetic expression of.

That is, the antecedent part membership function is calculated in the first layer 11 using y = 1 / (1 + exp (-ωx + θ)) which is the input / output relationship of the sigmoid neuron 2 as described above. Although the configuration is adopted, since the differential value of “y” is represented by ∂y / ∂x = y (1-y) ω, the antecedent output from the linear neuron 1 of the first stage layer 11 is obtained. Using the function value M of the membership function and the internal state value ω that realizes the antecedent membership function,
By calculating the value of the arithmetic expression “M (1-M) ω”, the differential value of the antecedent membership function of each fuzzy rule is calculated.

Then, the output means 24 outputs the calculated differential value of the membership function of the antecedent part as consult information. The consult information output in this manner represents the rate of change between the input value input to the fuzzy structure type neurocomputer 100 and the output value output from the fuzzy structure type neurocomputer 100. According to this consult information,
It becomes possible to more accurately grasp the financial data items to be operated and the operation amount thereof in order to obtain the desired bond rating information, more accurately than ever before.

At this time, when the multiplication means 23 is provided, the multiplication means 23 calculates the differential value of the antecedent part membership function calculated by the calculation means 22 and the fuzzy rule from which the differential value is derived. Of the weight value (the internal state value between the second layer 3 and the third layer 13) is calculated, and the output means 24 replaces the differential value calculated by the calculation means 22 with The multiplied value will be output as consult information. The consult information output in this way becomes more accurate consult information by being weighted by the importance of the fuzzy rule.

On the other hand, the calculating means 22 replaces the differential value of the antecedent part membership function with the function value of the antecedent part membership function of the fuzzy rule output from the first layer 11 and the desired function value. A configuration for calculating a difference value from the value may be adopted. That is, when it is desired to obtain an output value that makes the consequent part membership function described by the fuzzy rule large, the function value of the antecedent part membership function of the fuzzy rule output by the first layer 11 and the desired value of that function value. The value is "1"
When you want to obtain the difference value from and the output value where the consequent part membership function described by the fuzzy rule becomes small,
The difference value between the function value of the antecedent part membership function of the fuzzy rule output from the first layer 11 and the desired value “0” of the function value is calculated.

Then, the output means 24 outputs the calculated difference value as consult information. The consult information output in this manner represents the magnitude of the deviation between the output value output by the fuzzy structure type neurocomputer 100 and the desired output of the output value. Accordingly, it becomes possible to more accurately grasp the financial data items to be operated and the operation amount thereof in order to obtain desired bond rating information, more accurately than ever before.

At this time, when the multiplication means 23 is provided, the multiplication means 23 calculates the difference value calculated by the calculation means 22 and the weight value (second value) indicated by the fuzzy rule from which the difference value is derived. An internal state value between the stage layer 12 and the third stage layer 13) is calculated and the output value is calculated.
Replaces the difference value calculated by the calculation means 22 and outputs this multiplication value as consult information. The consult information output in this way becomes more accurate consult information by being weighted by the importance of the fuzzy rule.

In FIG. 11, equity capital is 48.526 billion yen, ordinary income is 80.47 billion yen, and long-term debt composition ratio is 0.2455%.
A fuzzy structural neurocomputer 1 provides financial data of a company with an interest coverage ratio of 5.1830 times, an equity ratio of 0.3327, and bond rating information of BBB +.
An example of consult information output from the output means 24 when inputting to the bond rating determination device of 00 is shown in FIG. 12, and the equity capital is 202.483 billion yen, the ordinary profit is 31.843 billion yen, and the long-term debt is Fuzzy structural neurocomputer 10 provides financial data of a company with a composition ratio of 0.0243%, an interest coverage ratio of 32.35771 times, an equity ratio of 0.6413, and bond rating information of A +.
When inputting to the bond rating determination device of 0, output means 2
4 illustrates an example of consult information output from FIG. In the former case, the bond rating determination device of the fuzzy structural neurocomputer 100 outputs "0.368367", and in the latter case, the bond rating determination device of the fuzzy structural neurocomputer 100 outputs "0.528189". .

In the figure, the value of is the function value of the antecedent membership function output from the first layer 11, and the value of is
The value of the weight value indicated by the fuzzy rule is the differential value of the membership function of the antecedent part output as consult information,
The value of is the product of the differential value of the antecedent part membership function output as consult information and the weight value indicated by the fuzzy rule, and the value of is the function value of the antecedent part membership function output as consult information. And the difference value between the desired value of the function value and “0” is the difference between the function value of the antecedent part membership function output as consult information and the desired value of the function value of “1”. The difference value, is the product of the difference value output as the consult information and the weight value indicated by the fuzzy rule, and the value of the difference value is the difference value output as the consult information and the weight value indicated by the fuzzy rule. It is a multiplication value.

From the consult information shown in FIG. 11, in the case of the conventional technology, consulting, "to raise ordinary profit", can only be concluded in order to raise the rating of bonds. According to the invention, the consulting information concludes the consulting that "the ordinary profit should be raised or the interest coverage ratio should be raised." The consultant concluded that it should increase or increase the interest coverage ratio, and from the consult information of /, increase ordinary profit, increase equity capital, increase interest coverage ratio, long-term debt Concludes that consulting "Reduce the composition ratio"
According to the consultant's consulting information, the consultant concludes that "you should increase ordinary profit, increase equity capital, increase interest coverage ratio".

From the consult information shown in FIG. 12,
With conventional technology, in order to raise the bond rating,
While only consulting that "raise ordinary profit or raise interest coverage ratio" is concluded, according to the consulting information of the present invention, "whether to increase ordinary profit, The consultant concluded that the interest coverage ratio should be raised, and consulting information from the consultant concluded that "whether to increase ordinary profit, increase equity capital, interest
Based on the consultation information, the consultant concluded that "the coverage ratio should be increased," and from the consult information, "the ordinary profit should be increased, equity should be increased, or the long-term debt composition ratio should be decreased."
From the consult information of /, "Ordinary profit
The consulting, "You should increase your own capital" will be concluded.

Next, the financial consultation information collecting method of the present invention will be described in detail. FIG. 13 shows an embodiment of an apparatus configuration for realizing the financial consultation information collecting method of the present invention. In the figure, the same components as those described in FIGS. 1 and 2 are indicated by the same symbols.

In this embodiment, it is assumed that there are three bond rating agencies, each of which has different bond rating information. Correspondingly, three teacher signal managements are performed. The configuration includes the devices 200a-i (i = 1 to 3). The fuzzy structure type neuro computer 100 and the learning processing device 300a are also provided with three units corresponding to the teacher signal management devices 200a-i.

In such a configuration, when collecting financial consultation information for bond ratings, the user first needs to use a fuzzy rule that describes the correspondence between financial data and bond ratings. To generate. That is, a fuzzy rule such as the basic classification rule / modified classification rule described above is generated.

Next, the user applies the fuzzy rules generated in this way to the network configuration data processing means 20 of the three fuzzy structured neurocomputers 100.
Map to. This mapping process is executed according to the same procedure as described above.

Subsequently, the user selects three learning processing devices 3
00a is started. When activated in this way, each learning processing device 300a is associated with the corresponding teacher signal management device 200.
-When the financial data of the teacher signal managed by -i is input to the fuzzy structure-type neurocomputer 100, the conclusion value determination means 16 outputs a pair of bond rating values. Sigmoid neuron 2 of the first layer 11 according to
By learning the internal state values assigned to the internal connection to and the threshold values calculated by those sigmoid neurons 2, the function shape of the antecedent part membership function is tuned, and the linear shape of the third stage layer 13 is tuned. By learning the internal state value assigned to the inner connection to the neuron 1, the importance (weight value) of each fuzzy rule is tuned, and the inner connection to the linear neuron 1 of the fourth stage layer 14 is assigned. The function shape of the consequent membership function is tuned by learning the internal state value.

In this way, each learning processing device 300a
The fuzzy rules mapped to the network configuration data processing means 20 of the corresponding fuzzy structure type neurocomputer 100 will be automatically tuned to a more accurate one that matches the bond rating information of each bond rating agency.

When the learning process of the learning processing device 300a is completed, the user activates the financial consult information collecting device 400. When activated in this way, the financial consult information collection device 400 inputs the financial data to be financially consulted into the three fuzzy structural neurocomputers 100.

When this financial data is input, the output means 2 of the three fuzzy structure-type neurocomputers 100 is input.
4 outputs the calculated value calculated by each calculating means 22 or the multiplied value calculated by each multiplying means 23 to the outside as consult information, and thus the financial consult information collecting device 400 outputs this. Three types of consult information are collected as financial consult information, and are printed out, for example, in a format that facilitates comparison of these three types.

That is, the financial consult information collecting device 4
00, for example, has an equity capital of 48.526 billion yen and ordinary income of
Input the financial data of a company with 80.47 billion yen, long-term debt composition ratio of 0.2455%, interest coverage ratio of 5.1830 times, equity ratio of 0.3327 and bond rating information of BBB + to the fuzzy structural neurocomputer 100. When performing, one of the three fuzzy structural neurocomputers 100 outputs the financial consult information as shown in FIG. 11, and the other two fuzzy structural neurocomputers 100 follow their own conversion processing. Since such financial consult information is output, the three types of output consult information are collected as financial consult information.

The financial consult information thus collected represents the financial consult information regarding the bond ratings indicated by different bond rating agencies, and therefore the user refers to them and issues the corporate bond. In addition, the current rating is the same for both institution A and institution B, but if the ordinary profit is slightly increased, the institution B will be upgraded, so it is better to choose institution B. Can be executed accurately.

Moreover, the financial consult information collected in this way is such that, when the calculating means 22 adopts the configuration for calculating the above-mentioned differential value, the financial data input to the fuzzy structural type neurocomputer 100 and the fuzzy structural type information. The calculation means 2 represents the rate of change from the bond rating information output by the neurocomputer 100.
When 2 adopts the configuration for calculating the above-mentioned difference value, it represents the magnitude of the deviation between the bond rating information output by the fuzzy structural neurocomputer 100 and the desired output of the bond rating information. According to this financial consult information, it becomes possible to grasp the type of financial data to be operated and the amount of data operation to be operated in order to obtain the desired output more accurately than before.

Although the illustrated embodiment has been described, the present invention is not limited to this. For example, in the example,
Although the fuzzy structure-type neurocomputer 100 of the present invention has been disclosed according to the application example to the bond rating determination device, the present invention is not limited to this and can be applied to other data processing as it is.

Further, in the embodiment, the calculation means 22 uses the first
Although the present invention has been disclosed according to an example of using the function value of the fuzzy rule membership function output from the stage layer 11, the present invention is not limited to this, and the antecedent part of the fuzzy rule output from the first stage layer 11 is not limited to this. The function value of the membership function,
Since the grade value for the fuzzy rule output by the second layer 12 has a strong correlation, the function value of the antecedent membership function of the fuzzy rule output by the first layer 11 is replaced with the function value. , The grade value for the fuzzy rule output from the second layer 12 may be used.

[0104]

As described above, according to the present invention,
In the fuzzy structure-type neurocomputer that executes data processing according to the data conversion function of the network structure structured by the fuzzy rule, since the user takes the configuration to output appropriate consult information, the user follows the consult information. It becomes possible to accurately grasp the input variable to be manipulated to obtain the desired output and the manipulated variable thereof.

Further, according to the present invention, it becomes possible to collect the financial consult information which can realize the appropriate financial consulting for the bond rating, so that the user can obtain the appropriate financial consult information in accordance with the financial consult information. You will be able to perform financial consulting.

[Brief description of drawings]

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

FIG. 2 is a principle configuration diagram of the present invention.

FIG. 3 is an explanatory diagram of generated fuzzy rules.

FIG. 4 is an explanatory diagram of connection of network configuration data processing means.

FIG. 5 is an explanatory diagram of a defined membership function.

FIG. 6 is an explanatory diagram of a membership function shape.

FIG. 7 is an explanatory diagram of internal state values / thresholds of the first layer before learning is executed.

FIG. 8 is an explanatory diagram of bond rating values.

FIG. 9 is an explanatory diagram of internal state values after execution of learning.

FIG. 10 is an explanatory diagram of internal state values / threshold values after learning is executed.

FIG. 11 is an explanatory diagram of output consult information.

FIG. 12 is an explanatory diagram of consult information that is output.

FIG. 13 is an example of a device configuration for implementing the present invention.

FIG. 14 is an explanatory diagram of a fuzzy structure-type neurocomputer.

FIG. 15 is an explanatory diagram of a fuzzy rule.

FIG. 16 is an explanatory diagram of an antecedent part membership function.

FIG. 17 is an explanatory diagram of input / output signal characteristics of a sigmoid neuron.

FIG. 18 is an explanatory diagram of a consequent part membership function.

[Explanation of symbols]

 1 linear neuron 2 sigmoid neuron 10 input layer 11 first stage layer 12 second stage layer 13 third stage layer 14 fourth stage layer 15 fifth stage layer 16 conclusion value determination means 20 network configuration data processing means 21 internal state value management Means 22 Calculation means 23 Multiplication means 24 Output means 100 Fuzzy structure type neurocomputer 200 Teacher signal management device 300 Learning processing device

Claims (7)

[Claims] 1. A network configuration data processing means having a data conversion function configured by a network structure and configured to convert the data conversion function according to an internal state value, and a network structure of the network configuration data processing means. In the fuzzy structural neurocomputer, the network is structured according to a fuzzy rule of a data processing target, and a value that realizes an input / output data relationship of the data processing target is set as the internal state value. The differential value of the antecedent part membership function is calculated from the function value of the antecedent part membership function of the fuzzy rule output by the configuration data processing means and the internal state value that realizes the antecedent part membership function. Calculation means (22)
And a output means (24) for outputting the differential value calculated by the calculation means (22) as consult information, the fuzzy structure-type neurocomputer. 2. A network configuration data processing means having a data conversion function configured by a network structure and configured to convert the data conversion function according to an internal state value, and a network structure of the network configuration data processing means. In the fuzzy structural neurocomputer, the network is structured according to a fuzzy rule of a data processing target, and a value that realizes an input / output data relationship of the data processing target is set as the internal state value. The differential value of the antecedent part membership function is calculated from the function value of the antecedent part membership function of the fuzzy rule output by the configuration data processing means and the internal state value that realizes the antecedent part membership function. Calculation means (22)
And a multiplication means (23) for calculating a multiplication value of the differential value calculated by the calculation means (22) and the internal state value which will display the weight value of the fuzzy rule from which the differential value is derived, A fuzzy structure-type neurocomputer, comprising: output means (24) for outputting the multiplication value calculated by the multiplication means (23) as consult information. 3. A network structure data processing means having a data conversion function configured by a network structure and configured to convert the data conversion function according to an internal state value, and a network structure of the network structure data processing means. In the fuzzy structural neurocomputer, the network is structured according to a fuzzy rule of a data processing target, and a value that realizes an input / output data relationship of the data processing target is set as the internal state value. Calculation means (22) for calculating the difference value between the function value of the antecedent part membership function of the fuzzy rule output by the configuration data processing means and the desired value of the function value, and the calculation means (22) An output means (24) for outputting the difference value as consult information. Aji structural neurocomputer. 4. A network configuration data processing means having a data conversion function configured by a network structure and configured to convert the data conversion function according to an internal state value, and a network structure of the network configuration data processing means. In the fuzzy structural neurocomputer, the network is structured according to a fuzzy rule of a data processing target, and a value that realizes an input / output data relationship of the data processing target is set as the internal state value. Calculation means (22) for calculating the difference value between the function value of the antecedent part membership function of the fuzzy rule output by the configuration data processing means and the desired value of the function value, and the calculation means (22) The difference value and the weight value of the fuzzy rule from which the difference value is derived are displayed. A fuzzy structure type characterized by comprising: a multiplication means (23) for calculating a multiplication value with a state value; and an output means (24) for outputting the multiplication value calculated by the multiplication means (23) as consult information. Neuro computer. 5. The fuzzy structure-type neurocomputer according to any one of claims 1 to 4, wherein the calculating means (22) replaces the function value of the antecedent part membership function of the fuzzy rule output by the network configuration data processing means, A fuzzy structural neurocomputer characterized by performing a calculation process using a grade value for a fuzzy rule output from a network configuration data processing means. 6. A financial consult information collecting method for collecting financial consult information required for financial consulting for bond rating, wherein one fuzzy structural neurocomputer according to claim 1 to 5 is prepared. , A first processing step of structuring the network structure of the fuzzy structure-type neurocomputer according to a fuzzy rule that describes the correspondence between financial data and bond ratings, and multiple types of correspondence data between financial data and bond ratings Prepared and input the financial data for each correspondence data type to the fuzzy structured neuro computer structured in the first processing step, and in response to this input, the fuzzy structured neuro The output information output from the computer is bond rating information or bonds that are associated with the financial data. The second processing step of learning the internal state value of the fuzzy structural neurocomputer and the correspondence data learned in the second processing step so as to match the rating information with the information that enables identification. The internal state values for each type are sequentially set in the fuzzy structural neurocomputer, and the third processing step of inputting financial data for financial consulting to the fuzzy structural neurocomputer in association with this setting, By collecting the consult information output from the fuzzy structure-type neurocomputer as financial consult information in response to the financial data input in the processing process of No. 3, each correspondence relationship required for financial consulting for bond rating. A fourth processing step for collecting financial consult information for each data type, Characteristic financial consult information collection method. 7. A financial consult information collecting method for collecting financial consult information required for financial consulting for bond rating, wherein the fuzzy structured neurocomputer according to claim 1 is used for financial data and bonds. A first processing step for preparing the number of types of correspondence relationship data with the rating and structuring the network structure of each fuzzy structural neurocomputer according to a fuzzy rule describing the correspondence relationship between financial data and bond ratings, Financial data of the assigned correspondence data is input to each fuzzy structural neurocomputer structured in the first processing step, and in response to this input, the fuzzy structural neurocomputer outputs the financial data. Output information that is associated with the financial data, bond rating information or bond rating The second processing step of learning the internal state value of each fuzzy structure-type neurocomputer and the correspondence data learned in the second processing step so that the information matches the information that makes it possible to specify the information. The internal state value for each type is set in each fuzzy structural neurocomputer, and the third processing step of inputting financial data for financial consulting to each fuzzy structural neurocomputer and the third processing step By collecting the consult information output from each fuzzy structural neurocomputer in response to the financial data input as financial consult information,
A financial consult information collecting method, comprising: a fourth processing step of collecting financial consult information for each corresponding relationship data type necessary for financial consulting for bond rating.