JPH0424737A - Learning method for prewired neuro of rule part - Google Patents

Abstract

PURPOSE:To learn efficiently by learning the entire prewired neuro of a rule part after initializing a weight value for each of the antecedent implementing part of a membership function, a rule part and the consequent implementing part of a membership function by using a preliminarily stored knowledge or a random number. CONSTITUTION:An initial weight value is set for each of an antecedent implementing part 10 of the membership function, a rule part 12 and a consequent implementing part 14 of the membership function either by introducing the preliminarily stored knowledge or by using the random number, or by using both of them. Then, an entire network is learned by using a preliminarily prepared system learning data. Thus, it is possible to learn efficiently the control which is suitable for an objective system.

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 the initial weight values for each of the networks by introducing pre-existing knowledge, using random numbers, or a combination thereof, the entire network is configured to learn using system learning data prepared in advance.

[Industrial Field of Application] The present invention relates to a rule section pre-wire learning method 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 presented to a hierarchical neural network to learn and perform adaptive processing. Let it happen. 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 2adeh 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, Maq+nL was introduced in 1974.
An1 was the first to use it to control a steam engine.
Various types of fuzzy control are performed.

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.

■Obtain ambiguous control rules such as ``If the temperature is high, reduce the fire'' held by experienced operators.

■The meaning of the words (propositions) in the control rule is quantified in the form of a membership function, and the control rule is described in a rule expression of the type ``rlF-THEN part''.

■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 rule is, for example, “I, is big and !2is small
It is written as l thenhis 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, “x, is bi
If the truth value of gJ = 0.8 "x2it smallJ = 0,5, then according to the antecedent operation that selects the minimum value, the truth value = 0.5 is set to the consequent part of the fuzzy control rule. Process it so that it is output as the applied value for.

(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+
Applied value of is blgJ = 0.5r T+ is b
When the applied value of igJ=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 application 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 the above 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,
Antecedent membership function realization unit 1 before rule unit 12
0 and/or the subsequent consequent membership function realization unit 14, some of the units are internally coupled according to the control rule without internally coupling all the units, and the input control state quantity (X
,,X2. ...X7) The target is a rule section pre-wired neuro that outputs l or a plurality of control operation amounts (Y) corresponding to 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] All of the antecedent membership function realization unit 10, rule unit 12, and consequent membership function realization unit 14 are trained using the training data prepared in advance. Construct a learning method with a processing process that;

Here, as a method for setting initial weight values in the first step, initial weight values created based on knowledge held in advance (indicated as "a priori knowledge" in the figure) are set.

Another method for setting the initial weight values in the first process is to set the initial weight values created using random numbers. Initial settings are made using knowledge, and other initial weight values are set using random numbers.

Furthermore, at the start of learning in the second process, a learning flag is set to determine whether or not learning is to be performed for each connection of the hierarchical network structure consisting of the antecedent membership function realization unit 10, the rule unit 12, and the consequent membership function realization unit 14. is set, and the weight value is optimized through learning processing based on the learning flag set for the connection.

[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, we need to know in advance the rule part prewire neuron, which has a hierarchical network configuration consisting of an antecedent membership function realization part, a rule part, and a consequent membership function realization part, which realize the fuzzy control rules of the target system. The initial weight values are set based on , or the initial weight values are set based on random numbers, or both are used in combination.

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 the pack propagation method is performed for the entire network, in which the initially set weight values are corrected so as to substantially match the signal).

By learning the entire rule section pre-wired neuron at the same time in this way, it is possible to perform learning that maximizes the performance of the network.

Furthermore, learning efficiency can be improved by initializing the weights of the antecedent membership function realization section, the rule section, and the consequent membership function realization section into which knowledge held in advance is introduced.

Furthermore, an antecedent membership function realization part using random numbers,
By initializing the weights of the rule part and the consequent membership function realization part, the performance of the network can be maximized.

[Embodiment] FIG. 2 is a learning operation flowchart showing a processing procedure as a first embodiment of a learning method for a rule section pre-wire neuro according to the present invention. The rule section pre-wire neuron that is the target of the learning method according to the procedure of FIG. 2 is, for example, the antecedent membership function realization section 10 and the rule section 12 shown in FIG.
, and a consequent membership function implementation section 14 (including a centroid calculation implementation section 18)) for realizing a fuzzy control rule. Here, the rule part prewired neuro does not internally connect all units between the rule part 12 and the antecedent membership function realization part 10 in the preceding stage and/or the consequent membership function realization part 14 in the latter stage. It is characterized by a network structure in which some parts are internally connected according to control rules, and is contrasted with a rule-part fully connected neuron in which all internal connections are connected.

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

[Step S11 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 Initial setting is performed by introducing knowledge held in advance so as to realize the fuzzy rule given the weights Ccd and Cde of the co-rule unit 12.

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

[Step S4] The weight Cjg of the center of gravity calculation implementation unit 18 is initialized so as to realize the center of gravity calculation. Incidentally, in FIG. 3, it is included in the consequent membership function realization unit 14.

[Step S5] A group correspondence table for each phase in which the learning process is performed (a table that specifies the group of neurons to be learned, that is, a hierarchy) and a connection correspondence table for each group (a table that indicates the group of connections to be input to the group of units) ) settings.

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

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

[Configuration of the rule section prewire neuron] The rule section prewire neuron to which the present invention is applied has a hierarchical network configuration that can realize, for example, the fuzzy control rules shown in FIG.

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 reference unit in the rule section pre-wire neuron of FIG. 3.

In FIG. 4, the basic unit 20 constitutes a multi-power single output system, and includes a multiplication section 22 that multiplies a plurality of inputs by the weight value of each internal connection, and an addition section that adds the multiplication results of the multiplication section 22. and a threshold processing section 26 that performs threshold processing on the output of the addition section 24.

Here, if the first layer is the h layer and the second layer is the i layer, then the multiplier 2
2 and the addition unit 24 executes the calculation of equation (1) below, and the threshold processing unit 26 executes the calculation of equation (2) below.

x p, = Σy, hw, h (1
) Vol= 1/ ((1+e-'X"-"')
(2) However, h: Unit number of the h layer i: Unit number of the i layer p: Pattern number of the input signal θ1: Threshold value of the first unit of the i layer W+h: Weight value of the internal connection between h - the i layer X□ :h
Sum of products Vph of inputs from each unit of the layer to the 1st unit of the i layer: Output y from the hth unit of the h layer for the input signal of the pth pattern, , : 1 of the 1st layer for the input signal of the pth pattern Here, the above equation (2) is known as a sigmoid function, and the threshold processing section 26 of the basic unit shown in the black circle in the rule section pre-wire neuron in FIG. Performs threshold processing according to a function.

In contrast, the threshold processing unit 26 yp + = X.

The linear function unit set as follows becomes the white circle linear function unit in FIG. However, the input section, which is the first unit of the antecedent membership function implementation section 10, is excluded. As the hardware constituting the basic unit shown in FIG. 4, "Network configuration data processing device" disclosed in Japanese Patent Application No. 63-216865 can be used.

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 processes each input variable X, , X2 . ...Xn is input, input variable X1
Output the grade value of the membership function of ~Xn. 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 X, &X2 (SA) with variables X1 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 X1(S A
) to X n (LA) as input, and outputs the grade value of each rule.

Further, the consequent part rule part 12-2 contains the grade value L of the rule.
H8-1 to LH8-n are input, and grade values of membership functions y (SA) to y (LA) of output variables are output.

Further, the consequent membership function realization unit 14 (including the centroid determination unit) calculates the membership function y (SA) of the output variable.
Input the grade value of y (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 described as one neuron, but if the function of the unit can be specified, it may be a gate circuit, an arithmetic unit, etc. that realizes the unit 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, X, , 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 XI and X2 into two and supplies them to the antecedent membership function realization section 10. The antecedent membership function realization unit 10 includes four sigmoid function units 30-1 to 30.
-4 and four linear function units 32-1 inputting the outputs of each sigmoid function unit 30-1 to 30-4.
~32-3.

The rule section 12 includes sigmoid function units 34-1 to 34-8 that realize eight fuzzy rules, and the antecedent membership function realization section 10 for each sigmoid function unit 34-1 to 34-8 implements each fuzzy rule. Wiring is done according to the description. The outputs of the sigmoid function units 34-1 to 34-8 are outputted to the linear function units 36-1 to 36-3, which 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 generates membership functions based on two types of output values (grade values) output from the rule unit 12. is reduced (or enlarged) and the function sum of the two reduced membership functions is calculated.

Furthermore, the center of gravity calculation realization section 18 is composed of a center of gravity determining element output device 40 and a center of gravity calculation device 42, and the weighting of the output from the consequent part membership function realizing section 14 is provided in the center of gravity determining element output device 40. two linear function units 40-1
．． 40-2 to calculate the two centroid calculation elements Y
, , Y2 are determined, and finally the center of gravity is calculated from the two center of gravity calculation elements Y, , Y2 by the center of gravity calculation device 42 to generate the control output Y.

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

In Fig. 5B, there are seven neuron groups Ga-Gg, among which neuron group G in the input section
a and 2 of neuron group Gg of center of gravity calculation realization unit 18
Since these are merely separation and addition/synthesis, they are excluded from the learning targets of the present invention.
Therefore, the neuron group included in the learning target of the present invention is G
There are five: b, Gc, Gd, Ge, and Gf. On the other hand, the connection groups connecting each two knee long groups are Cab to C1.
Since learning is performed to correct the weight value of the input connection located in the previous stage of the neuron group to be learned by the Tsuk-Tsuku propagation method, the connection group Cab-Cel is subject to the learning processing according to the present invention. subject to.

Regarding the learning process according to the operation flow of 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 as follows in the connection groups Cab and Cbc of the antecedent membership function realization unit shown in FIG. 5B. The weight of each connection is
Initialize to implement the given membership function.

FIG. 6 shows the antecedent membership function realization unit 10 of FIG. 5A.
The control human power X1 side is extracted and shown.

In FIG. 6, the sigmoid function unit 30-1.3
The output value y output from 0-2 is y=1/ (1+e-
(W't-111) (3)

In the initial setting of the weight value for the antecedent membership function realization unit 10 of the present invention, for example, the weight value W is set to the input connection of the sigmoid function unit h30-1, and - 20 is set, and θ=-10 is set as the threshold value θ in the sigmoid function unit 30-1.

Further, the input connection of the other sigmoid function unit 302 is set as the weight value W, and at the same time, the threshold value θ is set as θ=10 in the sigmoid function unit 30-2. Note that the linear function unit 32-1. 32-
The weight value of the input connection No. 2 is set to the initial weight value -1.

In this way, in the sigmoid function unit 30-1, W<Q
When the weight value and threshold value are initially set such that θ<0, the unit output value XI (SS) for the control output X1 shown in the characteristic curve 44 in FIG. 7, that is, as the control human power X increases, 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.

Furthermore, in the sigmoid function unit 30-2, W〉0,
By initially setting the weight value and threshold value such that θ>0, the unit output value x, (LA) for the control human power X1 shown in the characteristic curve 46 in FIG. ) can be realized to calculate the truth value of a membership function with a function shape in which the function becomes large.

The sigmoid function unit 30-1 shown in FIG.
The initial setting of the weight value and threshold for 30-2 is such that the desired antecedent membership function is also realized for the remaining sigmoid function units 3C1-2 and 30-4 in the antecedent membership function realizing section 10. Initial settings are made based on prior knowledge.

[Initial Setting of Rule Section] The initial setting of the weight value for the rule section 12 shown in FIG. 3 is based on the connection group Cc included in the rule section 12 shown in FIG. 5B.
The weight values of each connection belonging to d and Cde are initialized by introducing knowledge held in advance so as to realize a 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 FIG. 8, eight sigmoid function units 34-
1 to 34-8 (see FIG. 3) correspond to fuzzy rules 1 to L8, respectively, which are extracted and shown on the right side. Therefore, eight sigmoid function units 34-1 to 34-3
For each of 4-8, use fuzzy rules 1 to L8.
The linear function units 32-1 to 32-4 provided in the antecedent membership function realization unit 10 located at the previous stage according to the
Connections are being made from. Fuzzy rules are also applied to the two linear function units 36-1 and 36-2 provided following the eight sigmoid function units 34-1 to 34-8, and the consequent part (T HE N) of 1 to L8 is applied. Wiring is done according to the following.

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.

FIG. 9 shows one sigmoid function unit 34 provided in the rule section 12, in which the truth value X,
,
will occur.

The output value y output from the sigmoid function unit 34 in FIG. 9 is 5-Wx+ ・X+ +Wx, ・X2-〇 (4)
y=1/ (1+e-S) (5)
It is calculated as The unit output value y given by the above equation (5) is the weight value Wx of the input connection.

W! By arbitrarily setting 2 and the threshold value θ, an appropriate function shape can be realized.

Specifically, the sigmoid function units 34-1 to 34-1 in FIG.
The numerical value shown in the input connection for 34-8 is set as the initial weight value, and the sigmoid function units 34-1 to 34-3
The numerical value shown below 4-8 indicates the threshold value θ.

For example, in the sigmoid function unit 34-1 in which the fuzzy rule L1 is set, the following is set, and the sigmoid function unit 34-1 realizes the fuzzy rule L1 by setting these initial values.
will perform threshold processing of the function shape shown by the three-dimensional coordinates in FIG.

Hereinafter, similarly, sigmoid function units 34-2 to 34-
By setting initial values of weights and thresholds for each of the fuzzy rules 2 to L8 of 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 connection for -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 settings of the consequent membership function realizing unit] The initial setting of the weight values for the consequent membership function realizing unit 14 in FIG. 5A is such that the weight values of the connections belonging to the connection group Ce4 in FIG. 5B are set in advance. Initial settings are made to realize the given membership function based on the knowledge of 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 units 36-1 to 36-2 of the rule section 12 at the previous stage as shown in the figure. These nine units 38-1 to 38-9 correspond to points in the range O to 1 on the training line of the control output y at equal intervals.

The initial setting of the weight values for the input connections to the linear function units 38-1 to 38-9 is performed by controlling the horizontal axis shown on the right side for the connection weight values of the output y (SS) from the unit 36-1 of the rule section 12. The unit 38- according to the characteristic curve 46 of the graph with the output Y1 vertical axis as the unit output value.
From the 1 side to the weight value on the 38-9 side, 1, 1. 1.0.75.0.5.0.25. 0.
0. Initialize 0.

Conversely, for the connection from the linear function unit 36-2 of the rule section 12 that produces the unit output V (LA), the weight values 0, 0. 0.0.25.0.5.0.75. 1.
1. Initialize 1.

Furthermore, the linear function units 38-1 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 unit output value y(
SS) and y (LA) are reduced by multiplication by initial weight values, added together, and output to the next center-of-gravity 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 is also possible to set a weight value of 1 or more to enlarge it.

[Initial Setting of Center of Gravity Calculation Implementation Unit] The weight values for the center of gravity calculation implementation unit 18 shown in FIG. 5A are initialized so as to realize center of gravity calculation for the 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 two linear function units 40-1 and 40-2 provided in the centroid determining element output device 40, the nine linear function units 38-1 provided in the consequent membership function realization unit 14 of the preceding stage ~38-9 are connected. Then, for this input connection, as shown in the figure, two centroid derived values Y1. A weight value for calculating Y2 is set.

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

Y= f g'ade(y)ydy / f g+ad
e(y)dy (6) In order to realize this formula (6), the linear function unit 38 provided in the consequent part membership function realization unit 14 for the linear function unit 40-1
As weight values for the connections from -1 to 38-9, weight values that increase from 0 at equal intervals starting from the minimum value Y=0 of the control output Y and reaching the maximum value Y=1 are assigned.

On the other hand, for the linear function unit 40-2, the weight decreases from 0 to -1, for example, at the same equal intervals as above, from the maximum value Y = 1 of the control output Y to the minimum value Y = O. Assign a value.

By weighting the input connections to the linear function units 40-1 and 40-2, assuming that the uni-soto output values of the preceding linear function units 38-1 to 38-9 are C7 to C9, the linear function units 40- 1 is Yl=IC+ +0.875 C2+(], 75C3+
0.625 C4+0.5 C5+0J75 C
6+0.25C,+O,I25C,+OC9(7)
and the linear function unit 40-2 outputs Y1=O
C, -0, +25 C2-0, 25C30, 375C4
0,5C60,625C6-0,75C,-0,875
Output C8+IC9 (8).

The centroid derived values Y, Y2 output from the linear function unit 40-1, 40-2 are given to the centroid calculation device 42, and are calculated as follows: Y=Y2/(Y2 Yl) (
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 the equation (9) matches Y in the equation (6).

With the above processing, the initial setting of the weight values for the rule section wire neuron shown in S1 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 completion of the initial settings, a learning plan, ie, a learning schedule, for causing the rule section pre-wire neuro to perform learning processing is set. This learning schedule is set by setting the phase/group correspondence table shown in FIG. 20(a) and setting the group/connection correspondence table shown in FIG. 20(b).

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, after setting the initial weight values, the antecedent membership function realizing section 10. Rule section 12. Since the consequent membership function realizing unit 14 performs overall learning, as shown in FIG. 20(a), in phase 1, the neuron groups Gb, Gc, Gd, Ge shown in FIG. , ' Set five Gf.

On the other hand, the group/connection correspondence table sets the correspondence between the neuron groups Ga to Gg in the rule section prewire neuron and the input connection group, and is shown in FIG.
), the input connection groups Cab to Cgl are associated with the neuron groups Gb to Gg (see FIG. 5B).

When the setting of the phase/group correspondence table and the group/connection correspondence table is completed, proceed to S6 in Fig. 2, start the learning scheduler, and actually give the learning data to the rule section pre-wire neuron to perform the packing. The propagation method is used to perform learning to change the weight value of the long sword connection.

[Processing and Configuration of Learning Scheduler] FIG. 21 shows a flowchart of the learning process of the rule section pre-wire neuro of the present invention. This operation flow diagram is, for example, the 22nd
To explain the device configuration for learning processing that can be realized by the device configuration shown in the figure, the 22nd
In the figure, a learning instruction unit 50 and a learning unit 70 are provided for the roulet section pre-wire neuron 100.

The learning instruction unit 50 has a learning scheduler 52,
For the learning scheduler 52, there is a learning signal storage unit 54 that stores learning signals consisting of pairs of control inputs and desired control outputs for the control inputs, and a phase group storage unit 54 that stores the phase group correspondence table shown in FIG. A correspondence tape 56 and a group/connection correspondence table 58 storing a group/connection correspondence table are connected. Further, a learning convergence determination section 60 is provided for the learning scheduler 52.

On the other hand, the learning unit 70 is provided with a learning instruction reading section 72, a learning flag setting section 74, and a weight changing section 76 that changes 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 prewire neuron 100.

The learning adjuster 80 has the configuration shown in FIG.

Note that the weight change unit 76 is shown together with the learning adjuster 80.

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 includes a weight calculation unit 88 that performs weight calculation based on the pack propagation method, and a weight storage unit 9.
0 and a weight change amount storage section 92 are provided.

In this way, by providing the learning adjuster 80 for each connection of the rule section pre-wire neuro, the learning scheduler 52 of the learning instruction unit 50 can determine whether or not to change the weight value of the connection for the rule section pre-wire neuro 100. It is possible to set a learning flag to be determined, and it is possible to change the weight, that is, perform learning, only for connections for which the learning flag is valid, that is, the flag is turned on.

Furthermore, the flag set for the learning coordinator 80 sets the antecedent membership that can execute 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 consisting of the function realization section 10, the rule section 12, and the consequent membership function realization section 14.

Adjustment of weight values when actually learning by backpropagation with a hierarchical network that conforms to fuzzy rules configured is performed as follows.

In FIG. 23, the weight change information reading unit 84 of the learning adjuster 80 monitors whether the weight change amount is written by the weight calculation unit 88 in the weight change amount storage unit 92 of the weight change unit 76. . When the weight change information reading unit 84 detects that the weight has changed, it notifies the weight change amount adjustment unit 86 of the fact. 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. On the other hand, if the learning flag is off, the weight change amount adjustment unit 86 is stored in the weight change amount storage unit 92 of the weight change unit 76.
Set to zero to disable weight changes. The hardware constituting this learning scheduler is patent application No. 63-227.
No. 825, "Learning method for network configuration data processing device".

Next, with reference to the learning processing flow shown in FIG. 21, the learning processing of the rule section pre-wired neuron according to the device configuration shown in FIG. 22 will be described.

First, when the learning scheduler 52 of the learning instruction unit 50 is activated, it first sets the phase counter i, which indicates the progress of the learning phase, to i=1 in Sl, and proceeds to S2, where the phase counter i is stored in the phase/group correspondence table 56. The neuron group number corresponding to learning phase i=1 is read out with reference to the correspondence table. In the present invention, in phase 1, the antecedent membership function realizing section 10 of the rule section prewire neuron 100. All neuron groups Gb of the rule part 12 and the consequent part membership function realization part 14
-Gf is set.

Next, the process advances to S3, and the connection number corresponding to the neuron group number is read out with reference to the group/connection correspondence table 58.

Next, the process proceeds to S4, where the connection group number is output and the learning flag of the learning adjuster 80 provided to the connection belonging to the connection group number is turned on.

Specifically, the learning scheduler 52 of the learning instruction unit 50 outputs a connection group number to the learning instruction reading section 72 of the learning unit 70, and the learning instruction reading section 72 reads the connection group number and sends it to the learning flag setting section 74. give. The learning flag setting unit 74 selects a learning adjuster 80 provided in the connection belonging to the connection group number that received the learning instruction.
The learning flag is set by instructing the flag storage unit 82 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 desirable control output for the control human power X, and sends the control human power X to the rule section prewire neuro 100. The teacher signal d is given to the weight change section 76 and the learning convergence determination section 60. Then, the control output Y outputted from the rule section pre-wire neuro 100 by the learning human power The learning convergence determining section 60 determines that learning is complete when the error of the control output Y with respect to the teacher signal d becomes equal to or less than a specified value, and ends the learning process.

When the learning process ends based on the learning execution command in S5,
In S6, phase counter i is set to 1. increment by
Check the end of the learning phase at S7. Since the learning process of the present invention ends in phase 1,
The series of processing ends immediately.

[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 neuron network, which is composed of a type of node called a unit and internal connections with weights.

In Figure 24, the h layer is the input device, and the i layer is the intermediate layer.
layer, and the j layer as the output layer. Of course, there is only one i-layer as the intermediate layer, but the same is true even if there are a plurality of intermediate layers. As shown in FIG. 4, each unit includes a multiplier 2 that multiplies each of the plurality of inputs by the weight of each internal connection.
2, an addition section 24 that adds all the multiplication results of the multiplication section 22, and a threshold processing section 26 that performs threshold processing on the addition results and outputs one output. The operations performed in this unit are as described in (1) and (2) above for the basic unit 20.
The formula becomes In the pack propagation method, the weight value Wih and threshold value θi of the hierarchical network are adaptively and automatically adjusted and learned using error feedback. (1) (
As is clear from equation 2), it is necessary to adjust the weight value Wih and the threshold value θi at the same time, but this task is difficult as they interfere with each other. Therefore, in addition to the normal unit, the applicant of the present application provides a unit whose output value is always 1 in the h layer on the input side, and assigns a threshold value θi as a weight value to the output of this unit with an output value of 1. Then, by incorporating the threshold value θi into the weight value Wih, the threshold value θj
has proposed a pack propagation method that can handle the weight value as a weight value (Japanese Patent Application No. 333484/1984).

By doing this, the above equations (1) and (2) become xp1°
Σyphw+h (yp+=1/ (1+exp(xp+)) in IIJ
(12), and the threshold value θi does not appear.

By expanding from equations (11) and (12) above, we get the following (13)
) (14) is obtained.

X p+= 'F;'l DIWII
(13) Vp+=1/ (1+exp(xH)
) (14) where, j : unit number Wl of the j layer, : weight value X2 of the internal connection between the i-j layers, sum of products V of inputs from each unit of the first layer to the third unit of the j layer, i: I)th pattern C) Output from the third unit of the third layer for the input signal The weight changing unit 76 in FIG.
When receiving the teacher pattern d□, which is the signal to be taken by the output cover turn Vp+ and the output cover turn Y+++ output from the output layer j when δp+-V++ (1)'DI) (d.+5'
-1) Calculate (15). Subsequently, ΔW,, (+)=εΣδpjy++++ζΔWu(1-
1) However, ε: Learning constant ζ: Momentum t · Calculate the update amount ΔW(1) of the weight value between the i layer and the j layer according to the number of learning times. Here, the reason for adding the weight value update amount determined in the previous update cycle in the second term on the right side of equation (16) is to speed up learning.

Subsequently, the weight value changing unit 76 changes δ1. calculated using the above equation (15). using δ. +−Yp+(I Yp+)Σδ. +W+6(1-
1) Calculate (17), then ΔW+h(1) =εΣδplyDh+ζΔWlh(t
-1), update amount of weight values of h layer and i layer △w+h(
+) is calculated.

Subsequently, the weight change unit 76 determines the weight value for the next update cycle according to the update amounts calculated by the equations (16) and (18). The output pattern Vp+ for the input pattern pattern for learning the update of the above weight values is the teacher pattern d+.
The learning process is repeated so that a weight value matching +I is obtained.

On the other hand, if the units constituting the hierarchical network in FIG. 24 are linear units, the above equation (14) becomes V , , =
(20). The above equations (15) and (17) in this linear function unit are δ,, = d p+ 3' −+
(21) δ, l=photograph+W+1(t 1 )
(22).

[Weight Learning Process of the Present Invention] FIG. 25 is a processing flow diagram of weight learning performed by the pack propagation method for the rule section pre-wire neuron of the present invention shown in FIG.

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 in FIG. 5A. As shown, it has a six-layer structure from the first layer to the sixth layer.
The layer is made up of sigmoid function units, and the rest are made up of linear function units. Furthermore, the learning targets are the first to fifth layers, and the sixth layer in the final stage centroid calculation implementation unit 18 is excluded from the weight learning target.

Here, the fuzzy control value outputted by realizing the fuzzy rule by the rule part prewire neuron, that is, the control output value, is y6, and the learning process is performed from the 6th layer to the 5th layer, the 4th layer, the 3rd layer, and Perform the second layer, first layer, and so on sequentially. The progress of learning for each layer is indicated by an i counter. That is, in the initial state, the i counter is set to i=6. To further simplify the explanation, W, , , and ΔW1. I
-1 indicates a matrix, and the size is (number of neurons in the i-th layer) x (number of neurons in the i-1st layer).

Furthermore, δi and yi represent vectors whose size is the same as the number of neurons in the i-th layer.

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, in step 81 (hereinafter "step" will be omitted), a counter i for setting a learning target layer is initialized. Specifically, since there is a six-layer configuration as shown in FIG. 26, counter 1 is initially set to i=6. With this initial setting, the sixth layer is designated as the learning target.

Next, the process proceeds to 82, where it is determined whether the unit is a sigmoid function unit. In this case, since it is a linear function unit, S14
The process proceeds to step 815, where it is determined whether the layer is the final level, and since the sixth layer is the final level, the process proceeds to step 815, where the difference value δ6 is determined using the teacher signal d and control output y6 obtained at that time.

Next, the process proceeds to 86, where it is determined whether the unit is to learn weights. At this time, the 6th layer is excluded from the learning target, so
Specifically, since the learning flag of the learning adjuster 80 provided in the connection is off, the process advances to S8 and the weight value update amount ΔW6. Set ΔW65=0 and proceed to S9, in this case,
6w6. = 0, the weight value W65 is not updated. Next, proceed to S12 and check whether the counter i=1 or not.
That is, it is determined whether learning of all layers has been completed, and since it is i6 at this time, the process proceeds to 313, decrements the counter i by 1 to set j=5, and returns to S2 to start learning the next 5th layer. to go into.

Since the fifth layer is also a linear function unit, S2
The process proceeds to 814, but since it is not the final stage layer, the process proceeds to 316.
Proceed to step 6 and calculate the difference value δ6 obtained in the learning process of the previous 6 layers.
and weight value W6. Then, the difference value δ is determined according to a formula similar to the formula (22) above.

Next, proceed to 86 to determine whether or not it is a learning target,
Since the fifth layer is the consequent membership function realization unit 14, it is the learning target in the present invention, and therefore 8
7, and in S7, the difference value δ obtained in 816 is
6, the weight value update amount ΔW, 4 is calculated according to equations similar to equations (16) and (18) above. followed by S
Proceeding to step 9, the calculated weight update amount ΔW,4 is added to update the weight value W,4.

Similarly, the process is repeated for all units in the 5th layer, and when the process for the 5th layer is completed, the process proceeds to the 4th layer.

Since the fourth layer is also a linear function unit, the same process as for the fifth layer is repeated.

Further, when processing the third layer, since a sigmoid function unit is used in the third layer, 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 in S5. It is determined using the same arithmetic expression as the above-mentioned equation (17). Since this third layer is also a learning target, the process proceeds from 86 to 87, uses the difference value δ3 calculated in S5 to calculate the weight update amount ΔW32, and updates it in S9.

Repeat the same process up to the final unit of these three layers, and
Proceeding to the first layer, learning processing regarding linear function units is performed, and further advancing to the first layer, learning processing regarding sigmoid function units is repeated. Then, when the processing of the final unit of the first layer is completed and the process proceeds from S10 to 812, since the counter i is i=1 at this time, it is determined that the learning of all the layers has been completed, and a series of Finish the process.

[Weight Initial Setting 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 operation flow diagram of FIG.
L, S2. The present invention is characterized in that random numbers are used as the initial setting of weight values for the antecedent membership function realization unit JO, the rule unit 12, and the consequent membership function realization unit 14 of the rule unit pre-wire neuro in S3. Other 84°S5. The processing in S6 is the same as in the first embodiment shown in FIG.

Specifically, as shown in FIG. 6 for the antecedent part membership function realization part 10 and as shown in FIG. 8 for the rule part 12, the consequent membership function realization part As for No. 14, as shown in FIG. 18, the weights are randomly initialized using random numbers. Of course, for the sigmoid function unit, the weight value and threshold value of the input connection are set based on random numbers, and for the linear function unit, the weight value of the input connection is set using random numbers.

Furthermore, if there is some knowledge that is previously held, settings may be made based on this knowledge.

[Others] The above specific embodiment is as shown in FIG. 5A.
In this example, the pre-wired neuron rule section that realizes fuzzy rules for two control human forces be.

In addition, in the above embodiment, y"i/(1+e-wx") (2
3), as shown in FIG. 28, the unit outputs of the two sigmoid function units 94-1 and 94-2 are subtracted by the
may be given to realize the sigmoid function shown in FIG. The unit output y in this case is y = I/(1-e ″”x-” ) −1/(1-
It can be expressed as e−w2”e2).

[Effects of the Invention] As explained above, according to the present invention, the initial setting or the By training the entire rule section pre-wired neuron after initial setting using random numbers or a combination thereof, efficient learning that maximizes network performance is possible, and can be used in any fuzzy control adaptive system. It is possible to realize more efficient connection weight analysis for realizing fuzzy rules.

[Brief explanation of the drawing]

Fig. 1 is a diagram explaining the principle of the present invention; Fig. 2 is a learning operation flow diagram of the present invention; Fig. 3 is an embodiment configuration diagram showing the basic configuration of the rule section pre-wired neuron of the present invention; An explanatory diagram of the basic unit; Fig. 5A is a configuration diagram showing a specific embodiment of the rule section pre-wired neuron of the present invention; Fig. 5B is an explanatory diagram of the neuron group and connection group in Fig. 5A; An explanatory diagram of the initial settings of the antecedent part membership function realization part of the invention; FIG. 7 is an explanatory diagram of the membership function of the antecedent part; FIG. 8 is an explanatory diagram of the initial settings of the rule part of the invention; 10 is an explanatory diagram of the fuzzy rule L1 of FIG. 9; FIG. 11 is an explanatory diagram of the fuzzy rule L2 of FIG. 9
Explanatory diagram; Fig. 12 is an explanatory diagram of fuzzy rule L3 in Fig. 9; Fig. 13 is an explanatory diagram of fuzzy rule Upper 4 in Fig. 9;
4 is an explanatory diagram of fuzzy rule L5 in FIG. 9; FIG. 15 is an explanatory diagram of fuzzy rule L6 in FIG. 9; FIG. 16 is an explanatory diagram of fuzzy rule HIF in FIG. 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 group/connection correspondence table: Fig. 21 is a learning processing flow diagram of the present invention; Fig. 22 is a device configuration diagram of the learning processing according to the present invention; Fig. 23 is a diagram showing the configuration of the learning processing according to the present invention; Fig. 24 is an explanatory diagram of a general hierarchical neural network; Fig. 25 is a weight learning process flow diagram 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 1-6 Two-man power part 18: Center of gravity calculation realization part 20: Basic unit 22: Multiplication part 24: Addition Section 26 Two threshold processing sections 28-128-2.32-1 to 32-4.38-1 to 3
8-9.40-1.40-2: 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 Arithmetic unit 90, weight storage unit 92: weight change amount storage unit 96: subtraction unit

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) 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 according to the control rule without internally connecting all units. In a learning method of a rule section pre-wired neuro that outputs one or more control operation quantities (Y) corresponding to input control state quantities (X_1, X_2, ... X_n). , 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 the antecedent membership function realization unit (10), rule unit (12), and consequent membership function using pre-prepared learning data; A learning method for a rule section pre-wired neuro, comprising: a second step of learning all of the realization section (14); (2) The learning method of the rule unit pre-wired neuron according to claim 1, wherein the setting of the initial weight values in the first step is performed by setting initial weight values created based on knowledge held in advance. . (3) The learning method of the rule section pre-wired neuron according to claim 1, wherein the initial weight values in the first step are set using random numbers. (4) In the weight setting in the first step, weights that can be set based on knowledge held in advance are initially set using this knowledge, and other weight initial values are set using random numbers. How to learn the rules section pre-wired neuro. (5) At the start of learning in the second process, each connection of the hierarchical network consisting of the antecedent membership function realization unit (10), the rule unit (12), and the consequent membership function realization unit (14) 2. A learning method for a rule section pre-wired neuro according to claim 1, characterized in that a learning flag is set for determining whether or not learning is to be performed for the connection, and the weight value is optimized by a learning process based on the learning flag set for the connection. .