JPH0635889A - Neural network device and learning method therefor - Google Patents

Abstract

PURPOSE:To shorten a period of time required for relearning by corresponding to input/output characteristics in some part of an area in an input signal space by respective partial networks. CONSTITUTION:The input signal space of a neural network device is divided into plural input signal areas and a partial neural networks 103, 106 and 109 corresponding to the respective areas are provide. The partial neural network 103, 106 and 109 are made to initially learn only events belonging to the corresponding input signal areas and at the time of event updating, only the partial network corresponding to the area including the latest event is made to relearn. Then, at the time of every event updating, since only a part of the events is updated, the partial network requiring the relearning is extremely, limited. Since the respective partial networks only correspond to the input/output characteristics in a part of the area in the input signal space, network scale can be small, the number of the events belonging to the area in charge is smaller and the entire area is covered by a signal network.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network apparatus and a learning method for obtaining and outputting an output signal corresponding to an input signal given based on a correspondence learned from a case, and more particularly to the learning method. The present invention relates to a neural network device and a learning method thereof suitable for the case where it is necessary to continuously update a part of cases and re-learn during the use period of.

[0002]

2. Description of the Related Art In a neural network, when a large number of desired input / output cases are given and learned, the transfer characteristics of signals between neurons, which are constituent elements, change, and general correspondences included in the input / output cases. Acquire and reproduce relationships. This ability attracts attention as a learning machine,
Applications of neural networks are being tried to solve problems in various fields. If the characteristics of the target problem change with the passage of time, it is necessary to devise a method to adapt the input-output correspondence obtained through learning to this change. The usual general remedy is to update the set of cases to reflect changes in the characteristics of the problem, i.e. delete old cases, add new ones, and then retrain the network. Is. Proceedings of the IJNC N '9
0, Volume 1 (1990), pages 1 to 6 (Pro
c. IJCNN '90, Vol. 1 (199
0), pp. The neural network application example to the stock market forecast described in 1-6) is an example of using this method. In this application example, in order to adapt to changes in the economic environment and market conditions, only cases obtained from a certain length of time immediately before the forecast time are used for learning. Thus, each time we make a prediction, we add one newest case that is newly obtained and at the same time delete one of the oldest cases, keeping the set of cases reflecting the latest characteristics of the problem.

[0003]

The learning of the neural network by the sequential correction of the parameters requires a considerable amount of time in a practical problem, although it depends on the network scale and the number of cases. Even if only a small part of the cases is updated, the above-mentioned conventional technique re-learns the network using all the cases, so a lot of learning time is required for each update, and the update cycle is short. The problem was that it was not applicable. An object of the present invention is to provide a neural network device and a learning method therefor capable of re-learning in a short time when it is necessary to continuously update a part of cases and re-learn during a use period.

[0004]

In order to achieve the above object, the input signal space of a neural network device is divided into a plurality of input signal areas, and a partial neural network corresponding to each area is provided. The network is configured so that only the cases belonging to the corresponding input signal area are initially learned, and only the partial network corresponding to the area including the updated case is relearned at the time of updating the case. Furthermore, when re-learning is performed by sequentially correcting parameters, the parameter value immediately before re-learning is used as an initial value.

[0005]

With each case update, only a small number of cases are updated, so the number of sub-networks that require re-learning is very limited. Each sub-network only takes charge of the input / output characteristics of a certain area of the input signal space, so the network scale may be small, and
The number of cases that belong to the area in charge is small, and the time required for re-learning can be greatly shortened compared to covering all areas with a single network. In addition, depending on the problem, the convergence point where the parameters are sequentially modified does not shift much due to the update of the case.In such a case, if the value before re-learning is set to the initial value, it will converge with a small number of modifications, thus reducing the learning time. it can.

[0006]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 2 is a system configuration diagram of an embodiment of the present invention. Reference numeral 202 denotes a case to be given to a neural network (for example, when applied to a prediction problem of time series data, known actual values of explanatory variables and explained variables), input signal values (also actual values of explanatory variables), and various other types. It is an input device for inputting program input data and the like. Reference numeral 201 denotes a display device that displays an output signal value (a predicted value of an explained variable in the case of prediction) obtained from a neural network and output data of other various programs. Reference numeral 203 denotes a processing device that executes processing such as learning and application of a neural network and other various programs. Reference numeral 204 denotes a storage device that stores examples given to the neural network and various other programs and data. 20
Reference numerals 5, 206, and 207 denote a communication device, a communication line, and a data collection device for inputting cases and input signal values to the neural network online.

FIG. 1 is a block diagram of a neural network system which is an embodiment of the present invention. The device learns the desired input and output characteristics from the case, and when applying, calculates the output signal suitable for the input signal based on the learned correspondence relationship,
Output. The input signal space formed by the possible values of the input signal is divided into a plurality of areas called input signal areas. Hereinafter, the number of input signal areas is K. The learning input signal case group storage units 102, 105, 108, the partial neural networks 103, 106, 109, and the learning output signal case group storage units 104, 107, 110, etc. all have these K inputs. It is provided corresponding to each signal region. Further, the case is composed of a pair of an input signal value and an output signal value, and is classified by which input signal area the input signal value belongs to. The device is initially learned prior to application. The learning input signal case group storage units 102, 105, 108 and the like store the input signal value portions of the cases belonging to the corresponding input signal areas, and the learning output signal case group storage units 104, 1
07, 110, etc. also store the output signal value portion of the case. Partial neural networks 103, 106, 1
09 and the like are input signal case group storage unit 10 for learning.
2, 105, 108, etc. and learning output signal case group storage unit 1
Initial learning is performed using the corresponding cases stored in 04, 107, 110, etc. After the initial learning, the device applies the correspondence obtained in the learning to the input / output decision. The input signal classification processing unit 111 specifies the input signal region to which the input signal 114 given to the device belongs, and determines the partial neural network 10
A signal is supplied to the input end of the network corresponding to the specified area among 3, 106, 109 and the like. The output signal determination processing unit 112 responds to the partial neural network supplied with the signal by the input signal classification processing unit 111, based on the value obtained from the output end of the output signal 11 of the device.
Determine 5. The device updates the case on demand and is relearned to reflect the change in the I / O characteristics. When the input / output signal case classification update processing unit 101 is given an update request 113 regarding an input / output signal case such as addition, deletion, or correction, the input / output signal case classification update processing unit 101 identifies the input signal area to which the update-requested case belongs, and learns the input signal case. Group storage units 102, 105, 10
8 etc. and learning output signal case group storage units 104, 10
The storage unit corresponding to the specified area of 7, 110, etc. is updated according to the request. Partial neural network 1
03, 106, 109, and the like, respectively, corresponding learning input signal case group storage units 102, 105, 108, and
When the learning output signal case group storage units 104, 107, 110 and the like are updated, the updated cases are used to relearn.

FIG. 3 is a diagram for explaining a case where an input signal space (N dimensions) is divided into rectangular shapes as an example of the input signal area. It can be determined by whether or not the inequality of is satisfied.

[0009]

[Equation 1] However, among the vertices of the input signal area 301, When it is necessary to improve the continuity of the input / output characteristics obtained from the case near the area boundary, the input signal areas are arranged so that adjacent areas overlap each other. For example, in FIG. 3B, if the region adjacent to the region k having a larger coordinate value in the x _i direction of the coordinate is k + 1,

[0010]

[Equation 2] If so, an overlapping portion can be generated in both regions.

FIG. 4 is a diagram showing the structure of the input signal area definition table. This table is included in the input / output signal case classification update processing unit 101, and when the input signal classification processing unit 111 specifies the input signal area to which the input signal 114 belongs,
The input / output signal case classification update processing unit 101 requests the update request 113.
This is referred to when specifying the input signal area to which the case that has occurred exists. For the sake of concreteness of explanation, a rectangular input signal region is also taken as an example here. The input signal area definition table includes a rectangular area minimum coordinate value storage field 401, a rectangular area maximum coordinate value storage field 402, and an area identification number storage field 403. Each record registered in the table corresponds to one input signal area. In the rectangular area minimum coordinate value storage field 401,
The coordinate value of the minimum vertex of the input signal area is stored. In the rectangular area maximum coordinate value storage field 402, similarly, the coordinate value of the vertex with the maximum coordinate is stored. The area identification number storage field 403 stores a number for identifying an area. FIG. 5 is a diagram showing the structure of the learning input / output signal case table. This table is used for learning input signal case group storage units 102, 105,
108, etc. and learning output signal case group storage units 104, 10
7, 110, etc., which are made up of K partial tables 501, 505, etc. corresponding to the input signal area, when the partial neural networks 103, 106, 109, etc. perform initial learning and re-learning, It is referred to when the input / output signal case classification update processing unit 101 updates the case for which the request 113 is made. Each partial table of the learning input / output signal case table has case identification information storage fields 502 and 506.
Etc., input signal value storage columns 503, 507, etc., and output signal value storage columns 504, 508, etc. Each record registered in the partial table corresponds to one learning case belonging to the input signal area corresponding to the partial table. The case identification information storage fields 502, 506, etc. store information for identifying cases, such as time information in the case of time series data. The input signal value storage fields 503, 507, etc. store the input signal value of the case in the output signal value storage fields 504, 5
The output signal value of the case is stored in 08 and the like.

FIG. 6 is a flow chart showing an overall processing procedure of an embodiment of the present invention, and is a flow executed by each partial neural network. First, in step 601, the partial neural networks 103, 106, 109 and the like are initially learned using the corresponding cases, and the process proceeds to step 602. In step 602, the learned partial neural networks 103, 106, 109 and the like are applied to obtain and output the output signal 115 corresponding to the given input signal 114, and the process proceeds to step 603. In step 603, it is determined whether or not the condition for ending the operation of the apparatus is satisfied. If not satisfied, the process proceeds to step 604, and if satisfied, the present procedure is terminated. In step 604, it is determined whether or not there is a case for which updating is requested. If there is a case, the process proceeds to step 605, and if not, the process returns to step 602 and the above procedure is repeated thereafter. In step 605, the case is updated in response to the request, and the partial neural networks 103, 106, 109, etc. are relearned, and then the process returns to step 602 to repeat the above procedure.

FIG. 7 is a flow chart showing the procedure of the network initial learning process, which embodies the details of the process in step 601 of FIG. First, in step 701, an input / output signal case whose initial registration has not been processed is input, the input / output signal case classification update processing unit 101 is requested to add a case, and the process proceeds to step 702. In step 702, the input / output signal case classification update processing unit 101 compares the contents of the input signal area definition table in FIG. 4 with the input / output signal case requested to be added, and specifies the input signal area to which the case belongs, Proceed to 703. In step 703, the learning input signal case group storage units 102, 105, 108, etc., and the learning output signal case group storage units 104, 107, 11
0, etc. corresponding to the input signal area specified in step 702, more specifically, the input signal specified in the partial tables 501, 505, etc. of the learning input / output signal case table in FIG. Input / output signal examples are added to the partial table corresponding to the region, and the process proceeds to step 704.
In step 704, it is determined whether or not there is an input / output signal case for which initial registration has not been processed. If not, the process proceeds to step 705, and if there is, the process returns to step 701 to repeat the above procedure. In step 705, a network in which initial learning is not completed among the partial neural networks 103, 106, and 109 is stored in the corresponding learning input signal case group storage unit and learning output signal case group storage unit. Specifically, learning is performed using the case stored in the corresponding partial table of the learning input / output signal case table of FIG. 5, and the process proceeds to step 706. In step 706, it is determined whether or not there is a partial neural network for which initial learning has not been processed. If there is a partial neural network, the process returns to step 705 and the above procedure is repeated thereafter, and if not, this procedure ends.

FIG. 8 is a flow chart showing the procedure of the network application process, and is a detailed flow of the process in step 602 of FIG. First, step 80
In step 1, the input signal 114 is input to the device and supplied to the input signal classification processing unit 111. Next, in step 802, the input signal classification processing unit 111 compares the contents of the input signal region definition table of FIG. 4 with the input signal 114 to identify the input signal region to which the signal belongs. Next, in step 803, an input signal is supplied to the input terminal of the network corresponding to the input signal region specified in step 802 among the partial neural networks 103, 106, 109 and the like. Next, in step 804, the output signal determination processing unit 112 determines the output signal 115 based on the value obtained from the output terminal of the partial neural network in response to the input signal. Since the adjacent input signal areas are arranged in the input signal space so as to overlap each other, when values can be obtained from a plurality of partial neural networks, those values are obtained according to the degree of attribution of the input signal 114 to each input signal area. The output signal 115 is determined by dividing the values. Finally, in step 805, the output signal 115 is output from the device, and the procedure ends.

FIG. 9 is a flow chart showing the procedure of the network re-learning process, which embodies the details of the process of step 603 of FIG. First, step 901
Then, the unprocessed input / output signal case update request 113 is input and supplied to the input / output signal case classification update processing unit 101, and the process proceeds to step 902. In step 902, the input / output signal case classification update processing unit 101 compares the contents of the input signal area definition table in FIG. 4 with the input / output signal case for which the update request 113 is made, and specifies the input signal area to which the case belongs. Go to step 903. In step 903, the update content is determined, and if the case is added, the process proceeds to step 904, and if the case is deleted, the process proceeds to step 905. Although it is assumed here that case modification is represented by combining case deletion and addition, case modification can be provided as an independent function. In step 904, the storage unit corresponding to the input signal area specified in step 902 among the learning output signal case group storage units 104, 107, 110, and more specifically, the learning input / output signal case in FIG. The input / output signal case is added to the partial table corresponding to the specified input signal area of the partial tables 501, 505, etc. of the table, and the process proceeds to step 906. In step 905, the input / output signal case is deleted from the storage section or the partial table corresponding to step 904, and step 906 is executed.
Go to. In step 906, it is determined whether or not there is an unprocessed input / output signal case update request.
If so, the process returns to step 901 to repeat the above procedure. In step 907, a network for which re-learning has not been examined is selected from the partial neural networks 103, 106, 109, etc., and stored in the corresponding learning input signal case group storage unit and learning output signal case group storage unit. It is determined whether or not the case or the case stored in the corresponding partial table of the learning input / output signal case table in FIG. 5 is changed. If yes, the process proceeds to step 908, and if not, step 909. Go to. In step 908, the partial neural network in which the corresponding case has been changed is learned using the changed case, and the process proceeds to step 909. In step 909, it is determined whether or not there is a partial neural network for which re-learning has not been examined. If there is any, the procedure returns to step 907, and the above procedure is repeated. If not, this procedure ends.

10 to 12, details of the application and learning processes in the partial neural network will be described.
For specific explanation, a feedforward type layered network is used as the structure of the partial neural network, and an error backpropagation method is used as its learning method. FIG. 10 is a diagram for explaining a feedforward type layered neural network. In the feed-forward type layered network, neurons constituting the network are arranged so as to form multiple layers, and a signal applied from the outside is unidirectionally from the input side layer 1001 to the output side layer 1003. Connected to each other so as to transmit between neurons. The number of inputs, outputs, and layers of the network It is determined by the threshold value of the neuron, the input / output characteristic function, and the coupling strength between the neurons.

[0017]

[Equation 3]

[0018]

[Equation 4] As the input / output characteristic function of the neuron, a sigmoid type function like the following equation is often used.

[0019]

[Equation 5] FIG. 11 is a flowchart showing a procedure of a partial network application process. Is set to the value of the input signal to the network.

[0020]

[Equation 6] Next, in step 1102, the second layer 1002 is
To L-layer 1003

[0021]

[Equation 7]

[0022]

[Equation 8] Finally, in step 1103, the L-th layer 100 is calculated according to the following equation.
Output of each neuron of 3

[0023]

[Equation 9] FIG. 12 is a flowchart showing the procedure of the learning process of the partial network.

[0024]

[Equation 10]

[0025]

[Equation 11]

[0026]

[Equation 12]

[0027]

[Equation 13] Note that step 908 of the network re-learning process in FIG.
When this procedure is used for initialization, in order to shorten the processing time, it is possible to omit the initialization of the bond strength and the threshold and use the value before re-learning as it is as the initial value. In Step 1202, one case is selected from the learning case group of the network, and Step 1
Go to 203. This choice is simply taken in order In 203, the output signal value of each neuron of the first layer 1001 is calculated according to the equation (Equation 6). Then, in order from the second layer 1002 to the Lth layer 1003 according to the equations (Equation 7) and (Equation 8), Go to 05. In step 1205, the
Each neuron in layer 1003 Mu.

[0028]

[Equation 14] It In step 1206, the (L-1) th is calculated according to the following equation.
From layer 2 to layer 1002

[0029]

[Equation 15]

[0030]

[Equation 16]

[0031]

[Equation 17] However, η and α are parameters that determine the learning speed and stability. Then according to

[0032]

[Equation 18]

[0033]

[Formula 19] In step 1208, the learning end condition is determined. If the condition is not satisfied, the process returns to step 1202 and the above procedure is repeated thereafter. If the condition is satisfied, this procedure is terminated. As the learning end condition, for example, an error between the output signal and the case, the number of times the case is learned, or the like can be used.

FIG. 13 includes reconstruction of the input signal domain,
It is a flow chart which shows the whole processing procedure of other examples of the present invention. According to the present embodiment, the size of the input signal area can be always kept appropriate with respect to the number of belonging cases.
The procedure from steps 1301 to 1305 is the same as the procedure from steps 601 to 605 in FIG. After step 1305, the process proceeds to step 1306. In step 1306, the input signal region is reconstructed according to the number of cases, and the partial neural networks 103, 106, 10
9 and the like are relearned, and then the process returns to step 1302 to repeat the above procedure. Note that steps 1301 and 130
Details of the processing in steps 2 and 1305 are shown in FIGS.
Is the same as that shown in. FIG. 14 is a flowchart showing the procedure of the input signal area reconstructing process, which embodies the details of the process of step 1306 of FIG. First, in step 1401, the number of cases belonging to each input signal area is checked, and the input signal area having an excessive number of belonging cases is divided into a plurality of small areas. Next, at step 1402, the input signal area having the excessively small number of belonging cases is integrated with the adjacent area. Finally, in step 1403, the partial neural network is reconfigured so as to correspond to the input signal area after division / integration, and relearning is performed, and this procedure is ended. FIG. 15 is a flow chart showing the procedure of the input signal area division processing, and is the one in which the details of the processing in step 1401 of FIG. 14 are embodied. First, in step 1501, an input signal region that has not been considered for division processing is selected, and the flow advances to step 1502. In step 1502, the learning input signal case group storage units 102, 105, 108 and the like, and the learning output signal case group storage units 104, 107, 110 and the like corresponding to the input signal area selected in step 1501. Specifically, the number of cases registered in the partial table 501, 505, etc. of the learning input / output signal case table shown in FIG. 5 corresponding to the selected input signal area is compared with a predetermined reference value. If the number of cases exceeds the reference value, the process proceeds to step 1503, and if it is less than the reference value, the step 1
Proceed to 508. In step 1503, the size of the selected input signal region is compared with a predetermined reference value. If the region size exceeds the reference value, the process proceeds to step 1504, and if it is less than the reference value, the process proceeds to step 1508. In the case of a rectangular input signal area, for example, the length of the minimum side of the rectangle can be used as the area size, which is obtained from the contents of the input signal area definition table of FIG. Step 15
In 04, the axis dividing the area is determined, and step 150
Go to 5. For example, based on the contents of the input signal area definition table of FIG. 4 and the partial table of the learning input / output signal example table of FIG. 5, the input signal area is divided into two equal parts on a plane perpendicular to each coordinate axis direction of the input signal space. Then, choose the split axis that has the least deviation and divides the cases. Step 150
In step 5, the smaller of the number of cases after division according to the axis selected in step 1504 is compared with a predetermined reference value. If the number of cases exceeds the reference value, the process proceeds to step 1506, and the value is equal to or less than the reference value. If there is, go to step 1508. In step 1506, the input signal area is divided according to the axis selected in step 1504, and the newly obtained child area is set as shown in FIG.
Step 15 in the input signal area definition table of
Proceed to 07. In Step 1507, Step 1506
The cases are divided in accordance with the area division, and registered in a new partial table of the learning input / output signal case table of FIG. 5 created corresponding to the child areas, and the process proceeds to step 1508. In step 1508, it is determined whether or not there is an input signal region for which division processing has not been examined. If there is an input signal region, the process returns to step 1501 to repeat the above procedure, and if not, this procedure ends.
FIG. 16 is a flow chart showing the procedure of the input signal area integration processing, which embodies the details of the processing in step 1402 of FIG. First, in step 1601,
Select an input signal area that has not been examined for integration processing, and
Proceed to 602. In step 1602, the learning input signal case group storage units 102, 105, 108 and the like, and the learning output signal case group storage units 104, 107, 110 and the like corresponding to the input signal area selected in step 1601. Specifically, the number of cases registered in the partial table 501, 505, etc. of the learning input / output signal case table shown in FIG. 5 corresponding to the selected input signal area is compared with a predetermined reference value. If the number of cases is less than or equal to the reference value, the process proceeds to step 1603, and if it exceeds the reference value, the process proceeds to step 1605. In step 1603, the input signal area selected in step 1601 is integrated with the sibling area when the same area is generated as a child area by the input signal area division processing of FIG. 15, and the input signal area definition table of FIG. Restore the contents before splitting, and step 160
Go to 4. In step 1604, cases are combined according to the area integration in step 1603, the contents of the learning input / output signal case table of FIG. 5 are restored before division, and the process proceeds to step 1605. In step 1605, it is determined whether or not there is an input signal region for which integration processing has not been examined. If there is, the process returns to step 1601 and the above procedure is repeated thereafter, and if not, this procedure ends.

FIG. 17 is a flow chart showing the processing procedure of another embodiment of the present invention for time series prediction. The system configuration is the same as in FIG. Theory at a certain time t Value of the explained variable at time (t + Δt) after the prediction cycle Δt I do. First, in step 1701, the neural network device is initially learned using known real values of the explanatory variable and the explained variable as examples, and the process proceeds to step 1702. In step 1702, the current actual value of the explanatory variable is input, and the flow advances to step 1703. In step 1703, the device is supplied with the current actual value of the explanatory variable as an input signal, and in step 1
Proceed to 704. In step 1704, the output signal obtained from the device is output as a predicted value after Δt of the explained variable, and the process proceeds to step 1705. In step 1705, it is determined whether or not the condition for ending prediction is satisfied. If not satisfied, the process proceeds to step 1706, and if satisfied, the present procedure is terminated. In step 1706, wait for the prediction cycle Δt,
Proceed to step 1707. In step 1707, the actual value after Δt of the explained variable is input, and the flow advances to step 1708. In step 1708, the apparatus is requested to add the actual value of the explanatory variable obtained in step 1702 and the actual value of the explained variable obtained in step 1707 as the latest case, and the process proceeds to step 1709. In step 1709, the device is requested to delete the oldest case held by the device, and the process proceeds to step 1710. In step 1710, the device is relearned, then step 1702 is returned to and the above procedure is repeated.

[0036]

According to the present invention, in the neural network device and its learning method in which a part of the case needs to be continuously updated and re-learned during the use period, the time required for re-learning is shortened. There is. That is, in each case update, only a small number of cases are updated, so that the number of sub-networks requiring re-learning is very limited. Since each sub-network only handles the input / output characteristics of a certain area of the input signal space, the network scale can be small, and the number of cases belonging to the area in charge is small, and the entire area can be isolated. The time required for re-learning can be greatly shortened compared to the case where one network is used. In addition, depending on the problem, the convergence point where the parameters are sequentially modified does not shift much due to the update of the case.In such a case, if the value before re-learning is set to the initial value, it will converge with a small number of modifications, thus reducing the learning time. it can.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

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

FIG. 3 is a diagram illustrating an example of an input signal area.

FIG. 4 is a configuration diagram of an input signal area definition table.

FIG. 5 is a configuration diagram of a learning input / output signal case table.

FIG. 6 is a flowchart showing a processing procedure according to an embodiment of the present invention.

FIG. 7 is a flowchart showing the procedure of a network initial learning process.

FIG. 8 is a flowchart illustrating a procedure of network application processing.

FIG. 9 is a flowchart showing a procedure of a network re-learning process.

FIG. 10 is a diagram illustrating a feedforward layered neural network.

FIG. 11 is a flowchart illustrating a procedure of a partial network application process.

FIG. 12 is a flowchart showing a procedure of a learning process of a partial network.

FIG. 13 is a flowchart showing a processing procedure of an embodiment of the present invention including input signal area reconstruction.

FIG. 14 is a flowchart showing a procedure of input signal area reconstruction processing.

FIG. 15 is a flowchart showing a procedure of input signal area division processing.

FIG. 16 is a flowchart showing a procedure of input signal area integration processing.

FIG. 17 is a flowchart showing a processing procedure of an embodiment of the present invention for time series prediction.

[Explanation of symbols]

201 display device 202 input device 203 processing device 204 storage device 205 communication device 206 communication line 207 data collecting device 301 rectangular input signal area 401 rectangular input signal area definition table rectangular area minimum coordinate value storage field 402 input signal area definition table rectangular area Maximum coordinate value storage field 403 Input signal area definition table area identification number storage field 502 Learning input / output signal case table case identification information storage field 503 Learning input / output signal case input signal value storage field 504 Learning input / output Output signal value storage column of signal case table 506 Case identification information storage column of learning input / output signal case table 507 Input signal value storage column of learning input / output signal case table 508 Output signal value storage of learning input / output signal case table Column

Claims (9)

[Claims] 1. A neural network device for obtaining and outputting an output signal corresponding to an input signal given on the basis of a correspondence relationship learned from a case, wherein the plurality of input signals respectively correspond to a part of an input signal space. When a case storing means provided for each area and a case belonging to a corresponding input signal area are used to learn and a partial neural network and an input signal are given, the input signal area to which the input signal belongs is specified, When the input signal classifying means for supplying the input signal to the input terminal of the network corresponding to the specified region among the partial neural networks and the partial neural network supplied with the input signal by the input signal classifying means respond. The output signal determining means for determining the output signal based on the value obtained from the output end of the network is provided. And a neural network device. 2. The method according to claim 1, further comprising a case class updating means, and when receiving a case update request, an input signal area to which the case requested to be updated belongs is specified, and the case storage means is specified. The stored contents of the case storage means corresponding to the created area are updated in response to the request, and each partial neural network updates the stored contents of the corresponding case storage means. A neural network device characterized by re-learning using an updated case stored in a means. 3. The input signal region according to claim 1 or 2, wherein each input signal region is arranged in an input signal space so that adjacent regions have a common portion where they overlap each other, and a given input signal is the input signal region. When the regions belong to the overlapping part, the values obtained from the output terminals of a plurality of the network corresponding to the input signal region to which the signal belongs in each partial neural network are divided into output signals. Neural network device. 4. In any one of claims 1 to 3, when the cases stored corresponding to each input signal area are updated, the number of cases belonging to each area is checked, and the number of cases is excessive. The area is divided into a plurality of small areas, and areas with an excessively small number of cases are integrated with adjacent areas, and the memory contents and partial neural network of the cases are reconfigured to correspond to the areas after the configuration change. A neural network device characterized in that each partial neural network is relearned by using a corresponding case after being changed. 5. A learning method of a neural network device for obtaining and outputting an output signal corresponding to an input signal given on the basis of the correspondence learned from an example, wherein a plurality of learning methods respectively corresponding to a part of the input signal space are provided. A step of storing a case belonging to the input signal area for each of the input signal areas, a partial neural network is created for each of the input signal areas, and the case stored corresponding to the area is used,
In the step of performing initial learning and when an input signal is given, the input signal region to which the input signal belongs is specified, and the input signal is input to the input terminal of the network corresponding to the specified region among the partial neural networks. And the partial neural network responds to the input signal,
A learning method for a neural network device, comprising the step of determining an output signal based on a value obtained from an output end of the network. 6. In claim 5, when a case update request is received, the input signal area to which the case requested to be updated belongs is specified, and the case stored corresponding to the specified area is specified. Updating according to the request, using the updated example partial neural network corresponding to the updated example,
A learning method for a neural network device, comprising the step of re-learning. 7. The partial neural network according to claim 6, wherein the partial neural network learns a correspondence relationship between an input signal and an output signal of a case by sequentially adjusting the parameters of the partial neural network, and the relearning is performed at the time of relearning when updating the case. A learning method for a neural network device, characterized in that the adjustment is started with the value of the parameter immediately before as the initial value. 8. The method according to claim 6, wherein when the cases stored corresponding to each input signal area are updated, the number of cases belonging to each area is checked, and the area having an excessive number of cases is divided into a plurality of small areas. The step of dividing, the step of integrating the area where the number of cases is too small with the adjacent area, the memory contents of the cases and the partial neural network are reconfigured to correspond to the areas after the configuration change, and each changed part A learning method for a neural network device, comprising the step of re-learning using a corresponding case after changing a neural network. 9. A learning method of a neural network device for obtaining and outputting an output signal corresponding to an input signal given based on a correspondence learned from a case, wherein the input signal and the output signal of the device are respectively A step of configuring the predictive explanatory variables and the explained variables to correspond, storing known actual values of the explanatory variables and the explained variables as examples in each case storage means of the apparatus, and initially learning each partial neural network; And a step of giving an actual value of the explanatory variable newly obtained with the passage of time as an input signal to the input signal classification means of the neural network apparatus, and a predicted value of the explained variable corresponding to the actual value of the explanatory variable. , A step of obtaining an output signal from the output signal determining means of the neural network device, and an actual value of the explained variable corresponding to the actual value of the explanatory variable are obtained. Then, there is a step of requesting the case classification updating means to add both actual values as the latest case and delete the oldest case, and to predict time series data. Network device learning method.