JPH0424738A - Learning method for prewired neuro of rule part - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Summary] A control output corresponding to an input is generated according to a hierarchical network structure that realizes a fuzzy control rule, which is composed of an antecedent membership function implementation unit, a rule unit, and a consequent membership function implementation unit. Regarding the learning method of the rule part pre-wired neuro, we have developed an antecedent membership function realization part, a rule part and a consequent part membership function realization part, with the aim of efficiently learning to perform control suitable for the target system. After setting initial weight values for each of the following by introducing prior knowledge, using random numbers, or a combination of both, only the antecedent membership function realization part is learned using system training data prepared in advance. Configure it to do so.

[Industrial Application Field] The present invention relates to a learning method for a rule section prewired neuron that combines fuzzy control and a neural network.

In information processing using neurons, learning data consisting of a set of input patterns of the target system and desired output patterns for that input pattern is used in a hierarchical format.
It is presented to the physical network to learn and perform adaptive processing. In particular, a processing method called pack propagation method is attracting attention due to its high practicality in this learning process.

On the other hand, fuzzy theory was proposed by Zadeh in 1960, and is a theory that models the process of thinking and judgment based on ambiguity that humans make, such as whether the temperature is "low" or "high." This fuzzy theory is mathematically based on fuzzy set theory, which introduces membership functions to represent ambiguity.

As an application of fuzzy theory, in 1974 Maqmda
Various types of fuzzy control have been used since ni was first used to control steam engines.

Since the system that performs this fuzzy control handles uncertain things, it is necessary to have a configuration that can adaptively respond to adjustments and changes in the control algorithm.

[Prior Art] Generally, the following procedure is taken to construct a fuzzy control system.

■Experienced operators have vague control rules such as ``If the temperature is high, reduce the fire.''

■The meaning of the words (propositions) in the control rule is quantified in the form of a membership function, and the control rule is
Describe using a rule expression of the type “N~”.

■Conduct inspections through simulations and on-site tests.

■Evaluate inspection results and improve membership functions and rules.

To explain in more detail, the fuzzy control rules are, for example, rif 1. is big and x2is sm
It is written as all then7, is bigJ. Here, the IF section is called an antecedent section, and is a section that describes conditions regarding control state quantities (control inputs) such as temperature. Further, the THEN section is called a consequent section, and is a section that describes the conditions regarding the control operation amount (control output) of the operating terminal, etc.

Fuzzy control manages multiple fuzzy control rules prepared as the control logic of the target system, and also manages the meaning of ambiguous linguistic expressions such as "large" or "small" written in each fuzzy control rule. Quantify and manage it as a ship function.

The control procedure of the completed fuzzy control system is as follows.

■When a controlled state quantity such as temperature is input from the controlled object, the truth value is calculated according to the membership function of the antecedent part that has been set.

(2) Next, in accordance with the antecedent computation such as selecting the minimum value, a process is executed to determine the value to be applied to the consequent in each fuzzy control rule. For example, r x, ig b
Truth value of igJ = 0.8r l2is small
When the truth value of J = 0.5, the truth value = 0.5 is output as the applied value to the consequent part of the fuzzy control rule according to the antecedent part operation that selects the minimum value. to be processed.

(2) Next, a process is executed to determine an output application value from the input value (applied value) from the antecedent part to the membership function of the consequent part, which has the same amount of control operation.

For example, with two fuzzy control rules, r y+
Applicable value of is big J = 0.5 "Ft is big
When the applied value of J=0.6 is obtained, the applied value=0.6 is selected and determined according to the consequent operation that selects the maximum value.

■Subsequently, the membership function of the control operation amount is reduced (or expanded) according to the determined applied value, and the center of gravity of the figure of the function part of the membership function for the same control operation amount is calculated. The control operation amount, which is a fuzzy inference value, is calculated and output.

By the way, fuzzy control rules are planned to be generated according to the knowledge possessed by a skilled operator, and it is difficult to establish fuzzy control rules that can realize optimal control from the beginning. Therefore, it is necessary to improve the generated fuzzy control rules through trial and error while evaluating them through simulations and field tests.

Therefore, the inventors of the present invention developed a hierarchical network structure that realizes fuzzy control rules by applying neurology.
In other words, we have proposed an adaptive fuzzy control system using pre-wired neuro in the rule section (Patent Application No. 2-602).
No. 56 "Hierarchical Network Configuration Data Processing Device and Data Processing System").

That is, the rule part prewire neuro has a hierarchical network configuration that realizes a fuzzy control rule consisting of an antecedent part membership function realization part, a rule part, and a consequent part membership function.

Considering the automatic extraction of fuzzy control rules by the pre-wired neuron of the rule section, the following procedure may be taken.

■Set the membership function of the rule section prewire neuro.

■The rule section prewire neuro is presented with the input pattern of the target system and a set of desirable output patterns for that input pattern, and is made to learn.

■Extract the optimal fuzzy control rule for the target system by analyzing the weights of connections in the rule section.

[Problems to be Solved by the Invention] However, the learning performed for extracting fuzzy control rules for the rule section pre-wired neuron is
The learning method for learning, including the weights of the antecedent membership function part, rule part, and consequent part membership function realization part, has not been determined, and the performance of the roll part prewire neuron has not yet been determined. There is a need to develop efficient learning methods that can bring out the best in students.

The present invention has been made in view of such a situation, and an object of the present invention is to provide a learning method for a rule section pre-wired neuron that can efficiently perform learning for performing control suitable for a target system. .

[Means to solve the problem] FIG. 1 is a diagram explaining the principle of the present invention.

First, the present invention provides antecedent membership function realization unit 10.1
Or it has a hierarchical network structure consisting of a rule part 12 and a consequent part membership function realization part 14 which are made up of multiple hierarchies,
The rule part 12 is the previous antecedent part membership function realization part 1
0 and/or the subsequent consequent membership function realization unit 14, some of them are internally coupled according to the control rule without internally coupling all the units with
.

The target is a rule section pre-wired neuro that outputs one or more control operation amounts (Y) corresponding to X2+...X7).

In the present invention, as a learning method for such a rule part pre-wired neuron, [First step] Each of the antecedent part membership function realization part 10, the rule part 12, and the consequent part membership function realization part 14 is Setting initial weight values; [Second process] A learning method is configured that includes a process of learning only the antecedent membership function realizing unit 10 using training data prepared in advance.

Here, as a method of setting the initial weight values in the first process, the knowledge (in the figure Set the initial weight values created based on the information (shown as "a priori knowledge").

Another method for setting the initial weight values in the first process is to set the initial weight values created using random numbers in the antecedent membership function implementation unit 10, and then The implementation unit 14 is set with initial weight values created based on knowledge held in advance. Furthermore, as another method of setting the initial weight values of the antecedent membership function realization unit 10 in the first step, weights that can be set based on knowledge held in advance are initialized using this knowledge, and other weights are set using random numbers. You can also make initial settings.

Furthermore, at the start of learning in the second process, a learning flag provided for each connection of the antecedent membership function realization section 10 is turned on, and the learning flag of the rule section 12 and the consequent membership function realization section 14 is turned on. The learning flag provided for each connection is turned off, and only the weight values of the connections having the learning flag in the on state are optimized through learning processing.

[Operations] According to the learning method of the rule section pre-wired neuron of the present invention having such a configuration, the following operations can be obtained.

First, for the antecedent membership function execution part of the rule part prewire neuro, set the initial weight value based on the knowledge you have in advance, set the initial weight value using random numbers, or set the initial weight value using a combination of both. , initial weight values are set for the rule part and the consequent part membership function realization part based on knowledge held in advance.

Next, the learning process is started, the input cover turn for learning is presented to the rule section pre-wire neuron, and the output cover turn from the rule section pre-wire neuron is the learning amount cover turn (the teacher For example, learning using a pack propagation method is performed only on the antecedent membership realization unit, in which the initially set weight values are corrected so as to substantially match the signal).

In this way, by learning only the antecedent part membership function realization part of the rule part pre-wired neuron, we can realize the rule part and consequent part membership function with initial settings by introducing pre-existing knowledge. It is possible to adjust the weight of the antecedent membership function realization part while effectively utilizing the weight of the part.

In addition, learning efficiency can be improved by initializing the weights of the antecedent membership function realizing section into which pre-existing knowledge is introduced.

Furthermore, by initializing the weights of the antecedent membership function realization unit using random numbers, it is possible to maximize the neuro-based learning function.

[Embodiment] FIG. 2 is a learning operation flowchart showing the processing procedure in the first embodiment of the learning method for the rule section pre-wire neuro according to the present invention. The rule section pre-wired neuron that is the target of the learning process in FIG. Calculation realization unit 18
It has a hierarchical network structure for realizing fuzzy rules consisting of

Here, the rule section pre-wire neuro means the rule section 12.
has a network structure in which all the units between it and the antecedent membership function realization unit 10 in the previous stage and/or the consequent membership function realization unit 14 in the subsequent stage are internally coupled in accordance with the control rule. This is compared to a rule-part fully connected neuron that internally connects everything between hierarchical units.

In the learning process in FIG. 2, the following steps S1 to S6
Performs processing consisting of.

[Step SL] Weight Cab of the antecedent membership function realization unit 10.

Initial settings are made by introducing knowledge that is previously held so as to realize a membership function given Cbc.

[Step S2] Previous knowledge is introduced and initialized so as to realize the fuzzy rule given the weights Ccd and Cde of the rule section 12.

[Step S3] The weight Cel of the consequent membership function realization unit 14 is initialized by introducing knowledge held in advance so as to realize the given membership function.

[Step S4] The weight Cfg of the center of gravity calculation implementation unit 18 is initialized so as to realize the center of gravity calculation.

[Step S5] Create a neuron group correspondence table for each phase in which the learning process is performed (a table that specifies the neuron group to be learned, that is, a hierarchy) and a neuron group connection correspondence table (
(table showing the connection groups to be input to the unit group). That is, in the present invention, since only the antecedent membership function realization unit 10 is to be learned, the layers and specific details of the antecedent membership function realization unit are determined by the phase-group correspondence table. A neuron group is set for .

[Step S6] Start the learning scheduler and perform learning processing using the pack propagation method.

Hereinafter, steps 81 to 81 shown in the learning operation flow of FIG.
Each of S6 will be explained in detail.

[Structure of Rule Section Prewire Neuron] The rule section prewire neuron to which the present invention is applied has, for example, a hierarchical network structure capable of realizing the fuzzy rules shown in FIG. 3.

In Figure 3, the nodes shown in each part constitute one basic unit, and the basic unit consists of two functions: a sigmoid function unit shown by a black circle, and a linear function unit shown by a white circle.
seeds are used.

FIG. 4 shows the configuration of the basic unit in the rule section pre-wiring unit 2 in FIG. 3.

In FIG. 4, the basic unit 2o constitutes a multi-car output system, including a multiplication section 22 that multiplies a plurality of inputs by weight values of internal connections, and an addition section 24 that adds the multiplication results of the multiplication sections 22. and a threshold processing section 26 that performs threshold processing on the output of the addition section 24.

Here, assuming that the first layer is the h layer and the second layer is the i layer, the multiplication section 22 and the addition section 24 execute the calculation of the following equation (1), and the threshold processing section 26 executes the calculation of the following equation (2). Perform calculations.

Xn1=ΣVphW+h (
1) to V,, = 1 / (1+ e - 'Loto'')
(2) However, h: Unit number of the h layer i: Unit number of the i layer p: Input signal pattern number θ1 ・Threshold value W of the 1st unit of the i layer, , : Weight value of internal connection between h-i layer x, , : sum of products of inputs from each unit of h layer to the 1st unit of i layer y, h: 6 of h layer for input signal of p-th pattern
Output from No. 1 unit V-1: Output from No. 1 unit of i layer in response to input signal of p-th pattern. Here, the above equation (2) is known as a sigmoid function, and the rule section pre-wire in FIG. The threshold processing section 26 of the basic unit indicated by the black circle in the neuro section performs threshold processing according to the sigmoid function of equation (2).

On the other hand, a linear function unit in which 'ly I=X D is set in the threshold value processing section 26 becomes a linear function unit indicated by a white circle in FIG. However, the input section, which is the first unit of the antecedent membership function realization section 10, merely separates the control input into two, and does not need to function as a unit.

The hardware of the basic unit shown in Fig. 4 is a patent application filed in 1983.
No.-2685 entitled "Network Configuration Data Processing Apparatus".

Referring again to FIG. 3, the configuration of each part of the rule section pre-wire neuro will be explained.

First, the antecedent membership function realization unit 10 receives each input variable X+; In the figure, each membership function is a one-variable function for simplicity, but it may be a multi-variable function. In other words, it may be a membership function of Xs &X2 (SA) with variables Xn and X2 as inputs.

The rule section 12 includes an antecedent section rule section 12-1 and a consequent section rule section 12-2. The antecedent part rule part 12-1 is
Membership function X + (S
The grade values of A) to Xn(LA) are input, and the grade values of each rule are output.

Further, the consequent part rule part 12-2 contains the grade value L of the rule.
H3-1 to LH3-n are input, and grade values of membership functions y(SA) to y(LA) of output variables are output.

Furthermore, the consequent membership function realization unit 14 (including the centroid calculation realization unit 16) calculates the membership function V of the output variable.
Input the grade values of (SA) to V (LA) and output the value of the output variable Y.

Here, the antecedent part rule part 12-1 and the consequent part rule part 12
Since both the input and output of the rule unit 12 including -2 are the grade values of the membership function, these can be cascaded and multi-stage inference is possible.

Further, the neurons connecting each module 10.12-1.12-2.14 can be shared. For example, the output neuron of the antecedent membership function implementation unit 10 and the input neuron of the antecedent rule unit 12-1 may actually be the same.

Further, in FIG. 3, each unit is explained as one neuron, but if the function of the unit is specified, it may be a gate circuit, arithmetic unit, etc. that realizes a uni-soto function instead of a neuron.

FIG. 5A shows a specific embodiment of the rule section pre-wired neuron of the present invention, taking as an example a case where two inputs, Xn and X2, are given to obtain a control output Y.
The configuration of each part will be explained in detail with reference to the drawings as follows.

The input section 16 is composed of separation units 28-1 and 28-2, which divides the inputs Xn and X2 into two and supplies them to the antecedent membership function realization section 10.

The antecedent membership function realization unit 10 has four sigmoid function units 30-1 to 30-4, and four sigmoid function units 30-1 to 30-4, which input the outputs of each sigmoid function unit 30-1 to 30-4.
It is composed of linear function units 321 to 32-4.

The rule unit 12 includes sigmoid function units 34-1 to 34-8 that realize eight fuzzy rules, and the antecedent non-baship function realization unit 10 for each sigmoid function unit 34-1 to 34-8 outputs the sigmoid function units 34-1 to 34-8 for each fuzzy rule. Wiring is done according to the description. The outputs of the sigmoid function units 34-1 to 34-8 are linear function units 36-1 and 36 that perform consequent operations according to fuzzy rules.
2 are connected according to fuzzy rules.

The consequent membership function realization unit 14 is composed of nine linear function units) 38-1 to 38-9, and based on two types of output values (grade values) output from the rule unit 12, The membership function is reduced (or expanded) and the function sum of the two types of reduced (or expanded) membership functions is calculated.

Furthermore, the center of gravity calculation implementation unit 18 uses a center of gravity determining element output device 40.
and a center of gravity calculation device 42. The center-of-gravity determining element output device 40 weights the output from the consequent membership function realization unit 14 in each of the linear function units 40-1 and 40-2, and converts the sum to the center-of-gravity calculation element Y1. Find it as Y2. The center of gravity calculation device 42 also includes the center of gravity determining element output device 4.
The center of gravity is calculated using the two center of gravity calculation elements Y, , Y2 obtained in step o, and this center of gravity becomes the control output value Y as the fuzzy inference value.

FIG. 5B shows neuron groups and connection groups used to implement the learning method according to the present invention regarding the rule section pre-wire neuron of FIG. 5B.

In FIG. 5B, there are seven neuron groups (Ga-Gg) (seven layers), but among these, the neuron group Ga of the large sword section 16 and the neuron group Gg of the centroid calculation realization section 18 are simply separated and combined by addition. In the first place, learning the weights is not necessary because it is just a process.
The initialized weight values are used fixedly as they are.

On the other hand, the neuron group whose weight can be adjusted by learning using the pack propagation method is Gb-Gf(7)5
However, in the present invention, among these, neuron group Gb included in the antecedent membership function realization section,
The weight is adjusted using only Gc as the learning target.

On the other hand, the connection group connecting each two knee long loops is Ca
There are six b-Cfg, and learning is performed to adjust the weight values of the input connections located before the neuron group to be learned using the pack propagation method.
The connection group Cab-Cel becomes a learnable connection group. Of these, in the present invention, since only the antecedent membership function realizing section is the learning target, the weight values of the connection groups Cab and Cbc are adjusted by learning.

Regarding the learning process according to the learning operation flow shown in FIG. 2 for the rule section pre-wired neuron having the configuration shown in FIGS. 5A and 5B, the initial setting of weight values for each section will be explained first as follows.

[Initial settings of the antecedent membership function realization unit] The initial settings of the weight values for the antecedent membership function realization unit 10 are based on the connection groups Cab and belonging to the antecedent membership function realization unit shown in FIG. The weight value of each connection belonging to Cbc is set as follows to realize the given membership function.
Introduce the knowledge you already have and perform the initial settings.

FIG. 6 shows the antecedent membership function realization unit 10 of FIG. 5A.
The control human power Xn side is extracted and shown. In FIG. 6, the output value y output by the sigmoid function unit 30-1.30-2 is y=1/(1+e−”°”−
” ) (3).

In the initial setting of the weight value for the antecedent membership function realization unit 10 of the present invention, based on the knowledge that is available in advance, for example, the weight value W is set to the input connection of the sigmoid function unit 30-1, and W= −20 is set, and θ;−10 is set as the threshold value θ in the sigmoid function unit 3o-1. Further, W=20 is set as a weight value for the input connection of the other sigmoid function unit 30-2, and at the same time, θ=10 is set as a threshold value θ for the sigmoid function unit 3°2. Note that the linear function unit 30-1.
The weight value of the input connection 30-2 is set to the initial weight value=1.

In this way, W〈0
When the weight value and threshold value are initially set such that θ ku 0, the unit output value It is possible to set a truth value calculation function for a membership function having a function shape such that (SS) becomes small.

In addition, W〉0 in the sigmoid function unit 30-2
By initially setting the weight value and threshold value such that θ〉0, the unit output value Xn(LA) for the control human power Xn shown in the characteristic curve 46 in FIG. ) can be set to calculate the truth value of a membership function with a function shape that increases.

Sigmoid function unit 30-1.3 shown in FIG.
The initial setting of the weight value and threshold for 0-2 is such that the desired antecedent membership function is realized also for the remaining sigmoid function unit 30-3.304 in the antecedent membership function realization unit 10. Initialize weight values based on prior knowledge.

[Initial Setting of Rule Section] The initial setting of the weight value for the rule section 12 shown in FIG. 5A is based on the connection group Ccd belonging to the rule section 12 shown in FIG. 5B.
The weight value of each connection in Cde and Cde is initialized by introducing knowledge held in advance so as to realize the given fuzzy rule.

FIG. 8 is an explanatory diagram showing a specific example of initial setting of weight values for the rule section 12 together with fuzzy rules.

In Figure 8, eight sigmoid function units) 34-
1 to 34-8 correspond to fuzzy rules extracted and shown on the right side, and function shapes of membership functions shown in 1 to L8. Therefore, eight sigmoid function units 34-1~
For each of 34-8, use the fuzzy rule 1 to
Linear function unit 32 of the antecedent membership function realization section 10 located at the previous stage according to the description of the IF section in L8
-1 to 32-4 are connected. Also, 8
Two linear function units 36-1 to 36- provided following one sigmoid function unit 34-1 to 34-8
2, we also apply fuzzy rules to the consequent part of 1 to L8 (T
HEN section).

Here, the function shape of the fuzzy logic operation based on the weight values initially set for the input connections of the sigmoid function units 34-1 to 34-8 in the rule section 12 of FIG. 8 and the threshold values set for the units themselves will be explained as follows. become that way.

FIG. 9 shows one sigmoid function unit 34 provided in the rule section 12, in which the truth value
．． Assume that X2 is input and output y is generated by product-sum calculation and threshold processing.

The output value y output from the sigmoid function unit 34 in FIG.
It is calculated as 1/(1+e-s) (5). The unit output value y given by this equation (5)
is the weight value Wx1 of manual connection. By arbitrarily setting Wx2 and the threshold value θ, an appropriate function shape can be realized.

Specifically, the sigmoid function unit 34-1 in FIG.
The numerical value shown in the input connection for ~34-8 is set as the initial weight value, and the numerical value shown at the bottom of the unit is the threshold value θ.
shows.

For example, in the sigmoid function unit 34-1 to which the fuzzy rule L1 is set, are set, and the sigmoid function unit 34-1 that realizes the fuzzy rule L1 by setting these initial values is as shown in FIG. Threshold processing is performed on the function shape indicated by the three-dimensional coordinates.

Similarly, the remaining sigmoid function units 34-2 to 34-3
By initializing the weight values and threshold values for each of the fuzzy rules 2 to L8 in 4-8, the function shapes shown in FIGS. 11 to 17 can be obtained.

Note that the linear function unit 36 provided at the output stage in FIG.
The weight value of the input connection of -1.36-2 is initially set to 1, and the threshold value processing is a linear function, and the threshold value θ=0 is set.

[Initial setting of the consequent membership function realization unit] The initial setting of the weight value for the consequent membership function realization unit 14 in FIG. 5A is based on the weight value of the connection belonging to the connection group Cef in FIG. Based on the knowledge introduced, initial settings are made to realize the given membership function.

FIG. 18 is an explanatory diagram specifically showing the initial setting of weight values for the consequent membership function realization unit 14.

In FIG. 18, the consequent membership function realization unit 14
Nine linear function units 38-1 to 38- provided in
9, a connection input is performed from the linear function unit 36-1, 36-2 of the rule section 12 at the previous stage as shown in the figure. These nine hanging units 38-1 to 38-9 are
0 on the training line of the control output y in the order of 38-1 to 38-9.
They correspond to points in the range of ~1 at equal intervals.

The initial setting of the weight values of the input connections for the linear function units 38-1 to 38-9 is such that the connection weight value of the output y (SS) from one of the units 36 to 1 of the rule section 12 is controlled and output on the horizontal axis. According to the graph characteristic curve 46 with the Y1 vertical axis as the unit output value, the weight value is 1, 1 from the unit 38-1 side to the unit 38-9 side. 1.0.75.0.5, 0.25. 0
．． 0. Initialize to 0. Conversely, the unit output y (
A linear function unit 36- of the rule section 12 that generates
For the input connection from 2, weight values 0, O are applied from units 38-1 to 38-9 according to the characteristic curve
, 0, 0, 25, 0, 5, 0, 75, 1, 1, 1 are initialized. Furthermore, the linear function units 381 to 38-9 do not perform nonlinear threshold processing, and the threshold value θ is set to θ=0.

Due to the initial setting of the weight values in the consequent membership function realization unit 14, the linear function units 36-1.
The unit output values y (SS) and y (LA) outputted from 36-2 are reduced by multiplication by an initial weight value of 1 or less, then added, and output to the next centroid calculation implementation unit 18.

In FIG. 18, a weight value of 1 or less is initially set so as to reduce the unit output, but it may be expanded by initially setting a weight value of 1 or more.

[Initial Setting of Center of Gravity Calculation Implementation Unit] Initial settings of weight values for the center of gravity calculation implementation unit 18 shown in FIG. 5A are performed so as to realize center of gravity calculation for connection group C1g shown in FIG. 5B.

FIG. 19 shows a specific example of initial setting of weight values for realizing the center of gravity calculation performed on the center of gravity determining element output device 40 of the center of gravity calculation implementation unit 18.

For the 2-· linear function units 40-1 and 40-2 provided in the center-of-gravity determining element output device 40, the nine linear function units 38 provided in the consequent membership function realization unit 14 of the previous stage -1 to 389 are each connected. Then, as shown in the figure, weight values for calculating two gravity center derived values Y, , Y2 used in the gravity center calculation are set for this input connection.

In fuzzy control, there is a method of calculating the control output Y, which is a fuzzy inference value, by finding the center of gravity of a figure consisting of the sum of membership functions reduced for the same control output according to the following formula: adopt.

Y= f grade(y)ydy/f grade
e(y)dy (6) In order to realize this equation (6), the linear function unit 38- provided in the consequent membership function realization unit 14 for the linear function unit 40-1
As the weight values for the input connections from 1 to 38-9, weight values increasing from O at equal intervals starting from the minimum value Y=0 of the control output Y up to the maximum value Y=1 are assigned as shown in the figure. On the other hand, for the linear function unit 40-2, from the maximum value Y=1 of the control output Y to the minimum value Y=O, from 0 to the above-mentioned equal interval, for example, from O to -1
Assign decreasing weight values as shown.

By weighting the input connections to such linear function units 40-1 and 40-2, the linear function unit 3 in the previous stage is
If the unit output values of 8-1 to 38-9 are 01 to C9, the linear function unit 40-1 is as follows: Y1=I C+ + 0.875 C2+0.75C
3+0.625 C4+0.5 C5+[1,375C
60 o, 25C7+0.125 Cs + OC9(
7), and the linear function unit 40-2 calculates Y+
=OC+ 0.125 C20,25C30
,375C4-0,5C5-0,625C60,75C
Calculate 70,875CB + I C9 (8).

The centroid derived values Y, Y2 output from the linear function units 40-1, 4 (12) are given to the centroid calculation device 42, and are calculated as follows: Y=Y2/(Y2-Y+) (9
) is used to calculate the center of gravity, that is, the control output Y, which is a fuzzy inference value.

By substituting the above equations (7) and (8) into the above equation (9), it can be seen that (Y in equation 9 matches Y in equation (6)).

With the above processing, the initial setting of the weight value for the rule section pre-wire neuron shown in steps 81 to S4 in FIG. 2 is completed, and the process proceeds to step S5.

[Setting of learning schedule] In step S5 of FIG. 2, following the end of the initial setting, a learning plan, ie, a learning schedule, for performing learning processing on the rule section pre-wired neuron is set. The setting of this learning schedule is shown in Figure 20 (a
) and the settings of the phase/group correspondence table shown in (b).
) Perform the two steps shown in the group/wire connection correspondence table.

First, the phase-group correspondence table specifies the neuron groups to be learned in each phase as the learning phase progresses. In the present invention, since only the antecedent membership function realizing unit 10 is trained after setting the initial weight values, the 5B is set as a learning target group in phase 1, as shown in FIG. Neuron group Gb belonging to the antecedent membership function realization unit 10 shown in the figure,
Set up two GCs.

On the other hand, the group connection/correspondence table sets the correspondence between each of the two knee long groups Ga to Gg in the rule section prewire neuron and the input connection group, and is shown in FIG.
As shown in b), input connection groups Cab to C1g are associated with neuron groups Gb to Gg.

When the settings of the phase/group correspondence table and the group/connection correspondence table are completed, the process proceeds to step S6 in Fig. 2, where the learning scheduler is activated and the learning data is actually given to the rule section prewire neuron and packed. Learning is performed to adjust the weight values of input connections using the propagation method.

[Processing and Configuration of Learning Scheduler] FIG. 21 shows a flow diagram of the learning process performed by the learning scheduler of the rule section pre-wire neuro of the present invention. This processing flow can be realized, for example, by the device configuration shown in FIG. 22.

First, the device configuration for learning processing will be described. In FIG. 22, a learning instruction unit 50 and a learning unit 70 are provided for the rule section pre-wire neuron 100.

The learning instruction unit 50 has a learning scheduler 52, and a learning signal storage section 54, which stores a learning signal consisting of a control input and a desired control output corresponding to the control input, for the learning scheduler 52, as shown in FIG. A phase/group correspondence table 56 storing the phase/group correspondence table shown above, a group/connection correspondence table 58 storing the group/connection correspondence table, and a learning convergence determination unit 60
will be provided.

On the other hand, the learning unit 70 is provided with a learning technique reading section 72, a learning flag setting section 74, and a weight changing section 76 that changes and adjusts weight values using the pack propagation method.

Furthermore, a feature of the device configuration in FIG. 22 is that a learning adjuster 80 is provided in the connection connecting each unit in the rule section pre-wire neuron 100. The learning adjuster 80 has the configuration shown in FIG. In addition, learning adjuster 8
0 and the weight changing unit 76 are shown.

The learning adjuster 80 is provided with a flag storage section 82 , a weight change information reading section 84 , and a weight change amount adjustment section 86 . On the other hand, the weight change unit 76 is provided with a weight calculation unit 88 that performs weight calculation using the pack propagation method, a weight storage unit 90, and a weight change amount storage unit 92.

By providing the learning adjuster 80 for each connection of the rule section pre-wired neuron 100 in this way, 1. The learning scheduler 52 of the learning instruction unit can set a learning flag for the rule section prewire neuro 100 to determine whether or not to change the weight value of the connection, and the learning flag is valid, that is, the connection for which the flag is turned on. It is possible to change the weights, that is, perform learning, only for those.

In addition, the flag set for the learning coordinator 80 is set to implement an antecedent membership function that can implement fuzzy rules for a general hierarchical network that does not have a network structure that can implement fuzzy rules as shown in FIG. It is also used when specifying a hierarchical network structure consisting of a membership function realization part, a rule part, and a consequent part.

The weight values are adjusted as follows when actually learning is performed by pack propagation in a state where a hierarchical network conforming to fuzzy rules is configured.

In FIG. 23, the weight change information reading unit 84 of the learning adjuster 80 monitors whether the weight change amount is written from the weight calculation unit 88 to the weight change amount storage unit 92 of the weight change unit 76. When the weight change information reading section 84 detects a weight change, it notifies the weight change amount adjustment section 86 to that effect. The weight change amount adjustment unit 86 checks the learning flag in the flag storage unit 82, and does nothing if the learning flag is on. That is, the weight change amount calculated by the weight calculation section 88 is effectively written into the weight change amount storage section 92.

On the other hand, if the learning flag is off, the weight change amount adjustment unit 86
sets zero in the weight change amount storage section 92 of the weight change section 76, invalidating the weight change.

The hardware configuration of this learning scheduler was published in a patent application filed in 1983.
It is shown in "Learning method for network configuration data processing device" of No.-227825.

Next, referring to the learning process flow in Figure 21,
The learning process of the rule section pre-wired neuron according to the device configuration shown in the figure will be explained.

First, when the learning scheduler 52 of the learning instruction unit 50 is activated, step Sl (hereinafter [step J is omitted)]
Let the phase counter i indicating the progress of the learning phase be i=
It is set to 1 and the process proceeds to S2, where the neuron group number of the learning phase 11 is read out with reference to the correspondence table stored in the phase/group correspondence table 56. In the present invention, in phase 1, the rule section pre-wire neuro 1
Neuron groups Gb and Gc belonging to the antecedent membership function realizing unit 10 of 00 are set.

Next, proceed to S3, and set neuron group numbers Gb and Gc.
The corresponding connection numbers Cab and Cbc are read out with reference to the group/connection correspondence table 58.

Next, proceed to S4, and set the connection group numbers Cab, Cb.
c, and turns on the learning flag of the learning adjuster 80 provided for the connection belonging to the connection group number (, ah, Cbc. As a result, the rule section prewire neuro 10
The learning flags of the learning adjusters 80 provided on the connections of the antecedent membership function realization unit 10 of 0 are all turned on, while the learning adjusters 80 provided on the other connections
The learning flag for is flag-off.

Specifically, the learning scheduler 5 of the learning instruction unit 50
2 outputs the connection group numbers Cab and Cbc to the learning instruction reading section 72 of the learning unit 70, and the learning instruction reading section 72 outputs the connection group numbers Cab and Cbc.

Cbc is read and given to the learning flag setting section 74.

The learning flag setting unit 74 sets the learning flag by instructing the flag storage unit 82 of the learning adjuster 80 provided in the connections belonging to the connection group numbers Cah and Cbe that received the learning instruction to turn on the flag.

Next, the process proceeds to S5, where the learning scheduler 52 issues a learning execution command to start the learning process. This learning process reads the control human power X prepared in the learning signal storage section 54 of the learning instruction unit 50 and the teacher signal d that is a desired control output for the control human power X, and supplies the control human power X to the rule section prewire neuro 100. , the teacher signal is sent to the weight changing unit 7
6 and the learning convergence determination unit 60. And learning human power
The control output (fuzzy inference value) Y output from the rule unit prewire neuro 100 is taken into the weight change unit 76 and the learning convergence determination unit 60, and the weight change unit 76 changes the weight value according to the pack propagation method. The learning convergence determination unit 60 determines the end of learning when the change adjustment is made and the error of the control output Y with respect to the teacher signal d becomes equal to or less than a specified value, and notifies the learning scheduler 52, thereby ending the learning process.

When the learning process ends based on the learning execution command in S5,
Increment the phase counter i by one in S6, and
Check the end of the learning phase at 7. In the learning process of the present invention, since it ends at phase 1, the series of processes ends as is.

[Principle of learning by pack propagation method] Next, the principle of learning by the pack propagation method of the present invention will be explained together with the hierarchical network.

FIG. 24 shows a general hierarchical neural network, which is composed of a type of node called a unit and internal connections with weights.

FIG. 24 shows an h layer as an input layer, an i layer as an intermediate layer, and a j layer as an output layer. Here, the i-layer serving as the intermediate layer is one layer, but the same is true even if there are a plurality of intermediate layers. As shown in the basic unit of FIG. 4, each unit includes a multiplier 22 that multiplies each of the plurality of inputs by the weight of each internal connection, an adder 24 that adds all the multiplication results of the multiplier 22, and an adder 24 that adds all the multiplication results of the multiplier 22. It includes a threshold processing unit 26 that performs threshold processing on the result and outputs one output.

The calculations performed in this unit are the equations (1) and (2) shown above for the basic unit 20. In the pack propagation method, the weight value Wl of the hierarchical network
h and threshold value θ1 are adaptively and automatically adjusted and learned by error feedback. As is clear from equations (1) and (2) above, it is necessary to adjust the weight value W and the threshold value θ1 at the same time, but this task is difficult as they interfere with each other. Therefore, the applicant of this application provides a unit whose output value is always 1 in the h layer on the input side in addition to the normal unit, and assigns a threshold value θ as a weight value to the output of this unit with an output value of 1. Therefore, a pack propagation method has been proposed in which the threshold value θ1 can be incorporated into the weight value Wlh and the threshold value θ1 can be treated as a weight value (Japanese Patent Application No. 333484/1984).

By doing this, the above equations (1) and (2) become x,
1=Σy, hW lh (I
1. ) is yp+=1/ (1+exp(-x,+))
(12), and the threshold value θI does not appear.

By expanding the above equations (11) and (12), the following equations (13) and (14) are obtained.

Xn,=Σy,,W,1 (13)
yp+=1/ (1+exp(-xp+))
(14) However, j: unit number W of the j layer, ,: weight value x of internal connection between the i-j layers, ,: sum of products Vp+' of inputs from each unit of the i layer to the first unit of the j layer Output from the first unit of the j layer for the input signal of the p-th pattern The weight change unit 76 in FIG. When the teacher pattern apl, which is the signal that the cover turn Vpl should take, is received, the difference value between the two, (apr-yp+), is calculated, and then δp+=Yp+ (1-yp+) (apr-yp+
), (15) is calculated. Subsequently, ΔW++(+)=εΣδplVpl+ζΔWIl(t-
t) However, ε Learning constant ζ: Momentum t: Update amount of weight value between i layer and j layer according to the number of learning times ΔW++(f)
Calculate. Here, the reason for adding the second term on the right side of the equation (16) that is related to the update amount of the weight value determined during the entire update cycle is to speed up the learning.

Subsequently, the weight value changing unit 76 calculates δ calculated using the above equation (15).
, 1, and δ. ,=y, i (1-y++1) valley δ,,
Calculate W11 (+-]) (+7), then ΔWlh(1) =εΣδplVph+ζΔW+h(
According to 1-1), update amount ΔW+h of weight values of h layer and i layer
Calculate (1). Next, the weight value changing unit 76 changes the above (1
The weight value for the next update cycle is determined as follows according to the update amount calculated using equations (6) and (18). The above learning process is repeated so that the weight values are updated such that the output pattern y□ for the input pattern for learning matches the teacher pattern d□, and in fact, the error is less than or equal to the specified value.

On the other hand, if the units constituting the hierarchical network in FIG. 24 are linear function units, the above equation (14) becomes y p+= x pl (
20). The above equations (15) and (17) in this linear function unit are as follows: δ, I = d pj 'j pj
(21) δ, 1=Σδ−+Wn (t 1 )
(22).

[Weight Learning Process of the Present Invention] FIG. 25 is a process flow diagram of weight learning performed by the pack propagation method for the antecedent membership function realization unit 10 of the rule unit prewire neuro shown in FIG. 5A.

The parameters of each part in this learning process flow are the 26th
It is defined as shown in the figure.

FIG. 26 schematically shows the hierarchical structure of the rule section pre-wire neuron of FIG. 5A, and as shown, it has a six-layer structure from the first layer to the sixth layer. 1st layer and 3rd layer
The layer is made up of sigmoid function units, and the rest are made up of linear function units. Further, the learning targets are the first layer and the second layer belonging to the antecedent membership function realizing unit 10, and the remaining third to sixth layers are excluded from the weight learning target.

Here, the fuzzy inference value, ie, the control output value, output by the realization of the fuzzy rule by the pre-wired neuron in the rule section is set as y6, and the learning process is performed sequentially from the sixth layer at the final stage to the previous layer.

The progress of learning for each layer is indicated by an i counter.

That is, the i counter is set in the i layer 6 in the initial state.

The weight learning process flow shown in FIG. 25 will be explained as follows based on the parameters of each part shown in FIG. 26.

First, a counter i for setting a learning target layer is initialized using Sl. Specifically, since it has a six-layer structure as shown in FIG. 26, the counter j is initially set to the i layer 5. With this initial setting, the unit in the sixth layer is first designated as a learning target.

Next, the process proceeds to 82, where it is determined whether the learning designated unit is a sigmoid function unit. In this case, since it is a linear function unit, the process proceeds to S14, where it is determined whether or not it is the final layer.Since the sixth layer is the final level, the process proceeds to 815, where the teacher signal d and control output obtained at that time are A difference value δ6 is determined using y6.

Next, the process proceeds to 86, where it is determined whether the unit is to learn weights. At this time, since the 6th layer is excluded from the learning target, specifically, the learning flag of the learning adjuster 80 provided in the connection is off, so the process proceeds to S8 and the weight value change amount Δ
W6. Set ΔW65-0 and proceed to S9. In this case, ΔW
Since 65=0, the weight value W65 is not updated. Next, proceed to S12 and check whether the counter i is i=1 or not.
That is, it is determined whether or not the processing of the final learning layer has been completed, and since i=6 at this time, the process proceeds to 313 and the counter i is incremented by 1.
Decrement by one to set i=5, and return to S2 to start the next fifth
Enter layer learning.

Regarding the fifth layer, since it is a linear function unit, 8
Proceeds from 2 to 314, but since it is not the final level, it goes to 3.
16, a difference value δ5 is determined from the difference value δ6 obtained in the previous learning process of the sixth layer and the weight value W65 according to an equation similar to equation (22).

Next, proceed to 86 to determine whether it is a learning target,
Since the fifth layer belonging to the consequent membership function realizing unit 14 is excluded from the learning target, the process proceeds to S8 and sets the weight update amount ΔW, a=0, to invalidate the weight update in S9.

Similarly, the process is repeated for all units in the fifth layer, and when the process in the fifth layer is completed, the process proceeds to the fourth layer. Since the fourth layer is also a linear function unit, the same process as the fifth layer is repeated.

Further proceeding to the third layer processing, since the third layer uses a sigmoid function unit, the process proceeds from 82 to 83, and since it is not the final layer unit, the difference value δ3 unique to the sigmoid function unit is calculated at 85 (17 ) is calculated using an arithmetic expression similar to the formula. In the present invention of the third layer belonging to this rule section 12, since it is excluded from the learning target, the process proceeds to S8, the weight update amount ΔW3□=0, and the processing is repeated up to the final unit.

Next, when processing the second layer, since the second layer is a linear function unit, the process proceeds from 82 to 314, and since it is not the final stage, at 316, each difference value obtained in the processing up to the third layer is calculated. Calculate δ2 based on the weight value. Then proceed to S6,
In this case, the second layer is the antecedent membership function realization part]
Since it belongs to 0, it is a learning target in the present invention, and therefore the process proceeds to 87. In S7, the weight value update amount ΔW21 is calculated using the difference value δ2 obtained in S16 using equations similar to equations (16) and (18). Next, the process advances to S9 and the calculated weight value update amount ΔW2. By (
19) A new weight value W2. Calculate. Thereafter, the process is repeated for all units in the second layer, and when the process for the second layer is completed, the process proceeds to the first layer.

Since the first layer uses a sigmoid function unit, the process proceeds from S2 to 83, and since it is not the final layer unit, the difference value δ1 unique to the sigmoid function unit is determined in 85. Since this first layer is also a learning target, proceed from S6 to S7 and calculate the difference value δ in S5,
The weight update amount ΔWIO is calculated using , and updated in S9. The same process is repeated for the first layer, and when the process for the final unit is completed and the process proceeds from 810 to 312, the counter i at this time is i=1, indicating that learning for all layers has been completed. After making the determination, the series of processing ends.

[Initial Setting of Weights Using Random Numbers] Next, another embodiment of the learning method of the present invention will be described with reference to the operation flow diagram of FIG. 27.

In the learning process shown in the learning operation flow diagram of FIG. 27, random numbers are used as the initial setting of the weight value for the antecedent membership function realization unit 10 of the rule unit prewire neuro in step S]. It is characterized by The other processes in steps 84 to S6 are the same as in the first embodiment shown in FIG.

Specifically, for the antecedent membership function realizing unit 10, weight values are initially set randomly using random numbers as shown in FIG. In this case, the weight value and threshold value of the input connection are set based on random numbers. Furthermore, if there is some knowledge that is previously held, initial settings may be made based on this knowledge.

[Other Notes: In addition, as shown in FIG. 5A, the above embodiment takes as an example the rule section pre-wired neuron that realizes fuzzy rules targeting the two human control forces XI and X2. Of course, the number of inputs may be 1 or 3 or more as appropriate.

In addition, in the above embodiment, y=1/ (1+e-Wx'') (2
3), but as shown in Figure 28, two sigmoid function units)
94-1. Subtract unit output of 94-2 unit 96
may be given to realize the sigmoid function shown in FIG. In this case, the unit output y is y=1/(1 e "l x"' I ) I / (1 e
−w2 X−I+ 2 ).

[Effects of the Invention] As described above, according to the present invention, the antecedent membership function realization unit can initialize weight values based on knowledge it has in advance or initialize weight values using random numbers. ,
Furthermore, settings are made using both of them, and on the other hand, the weight values are initialized for the rule part and the consequent part membership function realization part based on the knowledge that each has in advance, and then the antecedent part membership function is realized. By learning only the part, it is possible to efficiently adjust the weights of the antecedent membership function realization part while making use of the weights of the rule part and the consequent membership function realization part, which have been initialized by introducing the knowledge that they have in advance. It is possible to perform a more efficient weight analysis of connections after an arbitrary control system is realized using fuzzy rules.

[Brief explanation of the drawing]

Fig. 1 is an explanatory diagram of the principle of the present invention; Fig. 2 is a learning operation flow diagram of the present invention; Fig. 3 is a configuration diagram of an embodiment of the rule section pre-wired neuron of the present invention; Fig. 4 is an explanatory diagram of the basic unit; FIG. 5A is a specific example configuration diagram of the rule section prewire neuron of the present invention; FIG. 5B is an explanatory diagram of the neuron group and connection group of FIG. 5A; FIG. 6 is an antecedent section membership of the present invention. FIG. 7 is an explanatory diagram of the membership function of the antecedent part; FIG. 8 is an explanatory diagram of the initial setting of the rule section of the present invention; FIG. 9 is an explanatory diagram of the rule section unit of the present invention; Figure 10 is an explanatory diagram of the fuzzy rule upper 1 in Figure 9; Figure 11 is an illustration of the fuzzy rule upper 2 in Figure 9.
Explanatory diagram: Fig. 12 is an explanatory diagram of the upper 3 fuzzy rules in Fig. 9; Fig. 13 is an explanatory diagram of fuzzy rule L4 in Fig. 9; 14th
Figure 15 is an explanatory diagram of fuzzy rule L6 in Figure 9; Figure 16 is an explanatory diagram of fuzzy rule H in Figure 9; Figure 17 is an explanatory diagram of fuzzy rule H in Figure 9. 8 explanatory diagrams on the rules: Fig. 18 is an explanatory diagram of the initial settings of the consequent membership function implementation unit of the present invention; Figure 19 is an explanatory diagram of the initial settings of the center of gravity calculation implementation unit of the invention; An explanatory diagram of the settings of the phase/group correspondence table and the group/connection correspondence table: Fig. 21 is a flow diagram of the learning process of the present invention: Fig. 22 is a diagram of the device configuration of the learning process according to the present invention; Fig. 23 is a diagram showing the configuration of the learning process of the present invention Fig. 24 is an explanatory diagram of a general hierarchical neural network; Fig. 25 is a flow diagram of the weight learning process of the present invention; Fig. 26 shows the parameters of each part in Fig. 25. Explanatory diagram; FIG. 27 is an operation flow diagram showing another embodiment of the present invention; FIG. 28 is an explanatory diagram of another membership function used in the present invention. In the figure, 10: Antecedent part membership function realization part 12: Rule part 14: Consequent part membership function realization part 16: Input part 18 Double center calculation realization part 20: Basic unit 22: Multiplication part 24: Addition part 26 Two threshold value processing units 28-128-232-1 to 32-438-1 to 3B-
940-1, 40-2: Linear function unit 30-1 to 30-234-1 to 34-8.94-1.9
4-2: Sigmoid function unit 40: Center of gravity determining element output device 42: Center of gravity calculation device 50: Learning instruction unit 52: Learning scheduler 54: Learning signal storage unit 56: Phase/group correspondence table 58 Niger loop/phase correspondence table 60: Learning Convergence determination section 70: Learning unit 72: Learning instruction reading section 74: Learning flag setting section 76: Weight changing section 80: Learning adjuster 82; Flag storage section 84: Weight change information reading section 86 Duplex change amount adjustment section 88: Weight calculation section 90: Weight storage section 92: Weight change amount storage section 96 Subtraction unit 100: Rule section Prewire neuro (a) All invention rules Fuzzy rule Ll: IFx[SS] O[SS) T) Correct Ny(S
S) L2: IFXl (S s) ANDX2 (SS)
THEN Y(LA)L3: rpm, A) ANDX2
(S S) THEN y(ss)L4:IFYl(L
A) ORxass) THEN Y(LA)L5:I
FXl(SS) oR5iz(LA)T HEN y<
ss>L6: Positive x l (S S) AND Co. (LA) π Positive NV (1-A) L7: IFXl (LA) ANDX7L
A) Force ENY (SS) L8: IFXKLA) ORX2
(LA) THEN Y (LA) (b) Ichiru Morijiro Phantom Concentration Diagram 98 Diagram 41 Detailed Rules for Awn Strike Unit Miko Myo Decoy Diagram 9 Phase: Group Correspondence Table Group: Connection Correspondence Table (b ) Figure 20 c'J(tc) I II II -11)-111 Ward C%, l('J q) ゝ---)-111-111-Y--- - Ward (-11-m-J C%J C'J Oko111--J CJC'-1I II II -1111110f)

Claims (5)

[Claims] (1) It has a hierarchical network consisting of an antecedent part membership function realization part (10), a rule part (12) consisting of one or more layers, and a consequent part membership function realization part (14), and the rule part (12) does not internally connect all the units between the antecedent membership function realization unit (10) in the previous stage and/or the consequent membership function realization unit (14) in the subsequent stage, but only some parts according to the control rule. In the learning method of one or more rule parts pre-wired neuro that is internally connected and outputs 1 or the control operation amount (Y) corresponding to the input control state quantities (X_1, X_2, ... X_n). and the antecedent membership function realization unit (10), the rule unit (12), and the consequent membership function realization unit (14).
a first step of setting initial weight values for each of; a second step of learning only the antecedent membership function realization unit (10) using training data prepared in advance; The learning method of the rule section pre-wired neuro is a feature. (2) The initial weight values in the first process are set by the antecedent membership function realization unit (10), the rule unit (12), and the initial weight values created based on knowledge held in advance.
, and the consequent membership function realization unit (14), respectively. (3) To set the initial weight values in the first process, initial weight values created using random numbers are set in the antecedent membership function realization unit (10), 2. A learning method for a rule part prewire neuron according to claim 1, characterized in that initial weight values created based on knowledge held in advance are set in the subject membership function realization part (14). (4) In setting the initial weight values of the antecedent membership function realization unit (10) in the first step, weights that can be set based on knowledge held in advance are initialized using this knowledge, and other weights are set using random numbers. 2. The method of learning a rule section pre-wired neuron according to claim 1, wherein the initial setting is performed using a pre-wired neuron. (5) At the start of learning in the second process, a learning flag provided for each connection of the antecedent part membership function realization part (10) is turned on, and the rule part (12) and the consequent part membership The rule according to claim 1, characterized in that a learning flag provided for each connection of the function realization unit (14) is turned off, and only the weight value of a connection having a learning flag in an on state is optimized by learning processing. How to learn part prewire neuro.