JPH0424736A - Learning method for prewired neuro of rule part - Google Patents

Abstract

PURPOSE:To adjust the optimal weight of the consequent implementing part of a membership function by learning the weight of the consequent implementing part of the membership function after learning the weight of either the antecedent implementing part of a membership function or a rule part, or the weight of both of them. CONSTITUTION:An initial weight value to which preliminarily stored knowledge is introduced 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. Then, the weight of either the antecedent part 10 or the rule part 12, or the weight of both of them is learned and finally the weight of the consequent part 14 is learned. The weight value can be initialized for the antecedent part 10 and the rule part 12 by using a random number or by using both the preliminarily stored knowledge and the random number. Thus, it is possible to adjust the optimal weight of the consequent implementing part of the membership function and to learn efficiently the control which is suitable to 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 value that introduces the knowledge held in advance for each, first, the weight of either or both of the membership function realization part of the antecedent part and the rule part is learned, and finally the weight of the consequent part is learned. The configuration is configured to learn the weights of the membership function realization unit. Note that the initial setting of the weight values of the antecedent membership function realization section and the rule section may be performed using random numbers, or may be performed using a combination of prior knowledge and random numbers.

[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, Maqmd+ was introduced in 1974.
+ni first used it to control steam engines,
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.

■Acquire ambiguous control rules such as ``If the temperature is high, reduce the fire'' that skilled operators have.

■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 ``rIF-THEN section''.

■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 I, is big and x2is sm
It is written as all theny, 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, is b
Truth value of igJ = 0.8r x2 ia 5ml1l
When the truth value of lJ = 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, r7.
Applicable value of is bigJ = 0.5 "7. 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 sum of the membership functions 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.
That is, we have proposed an adaptive fuzzy control system using prewired neuro in the rule section (Patent Application No. 2-602).
(See 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 controlling 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 that realizes a fuzzy control rule consisting of a multi-layered rule section 12 and a consequent membership function realization section 14, where the rule section 12 is an antecedent membership function realization section 10 in the preceding stage and/or a subsequent stage. The control state quantities (Xl, X2... Xゎ)
The target is a rule section prewire neuro 100 that outputs one or more control operation amounts (Y) corresponding to .

In the present invention, the following five learning methods are used as learning methods for the rule section pre-wired neuron 100.

■Claim 1; [First step antecedent membership function realization unit 10 has in advance knowledge (indicated as "prior knowledge" in the figure) or random numbers to initialize the weight values. At the same time, the rule section 12
and the weight values are initialized based on the knowledge that the consequent membership function realizing unit 14 has in advance;
Learn the weight of 0; [Third process] Consequent membership function realization unit 1 using learning data
The third process of learning the weights of 4 is configured as follows.

■Claim 2 [The first process corrule unit 12 initializes the weight values based on knowledge held in advance or using random numbers, and the antecedent membership function realization unit 10 and the consequent membership function realization unit 1
4. Initialize the weight values based on the knowledge held in advance; [Second process] Learning the weights of the rule section 12 using the learning data; [Third process] Using the learning data to set the consequent Part membership function realization part 1
Learn the weight of 4: Configure as follows.

■Claim 3 [The weight values are initialized based on the knowledge held in advance in the first process corrule unit 12 or using random numbers, and the antecedent membership function realization unit 10 and the consequent membership function realization unit 1
4. Initialize the weight values based on knowledge held in advance; [Second process: learn the weights of the rule section 12 using the learning data; [Third process] set the antecedent using the learning data. Part membership function realization part 1
When a weight of 0 is learned;
The fourth step is to learn the weight of 4.

■Claim 4 [First step] The weight values are initialized based on knowledge held in advance in the antecedent membership function realizing unit 10 or using random numbers, and the rule unit 12 and the consequent membership function are realized. Part 1
4. Initialize the weight values based on the knowledge held in advance;
Learn the weight of 0; [Third process] Use the learning data to learn the weight of the rule part 12; [Fourth process] Use the learning data to learn the consequent part membership function realization part 1
The weight of 4 is learned; it is configured as follows.

■Claim 5 [In the first step, the antecedent membership function realization unit 10 and the rule unit 12 initialize the weight values based on knowledge held in advance or by using random numbers, and realize the consequent membership function. Part 1
4. Initialize the weight values based on the knowledge held in advance; [Second process] Simultaneously learn the weights of the antecedent membership function realization unit 10 and the rule unit 12 using the learning data; [Third process] Consequent membership function realization unit 1 using learning data
The weight of 4 is learned; it is configured as follows.

■Claim 6 In the setting of initial weight values in the first process of the part where weights are first learned, weights that can be set based on knowledge held in advance are initialized using this knowledge, and other weights are initialized using random numbers. .

■Claim 7 At the start of learning of each process performed after the initial setting of the weight value of the first process, the antecedent membership function realization unit 10, the rule unit 1
2 and the consequent membership function realization unit 14, turn on the learning flag provided for each connection in the part that is the learning target, and turn on the learning flag provided for each connection in the part that is not the learning target. The flag 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, initial setting of weight values based on prior knowledge for the antecedent membership function realization part of the rule part prewire neuro and/or rule part, initial setting of weight values using random numbers, or a combination of both. The weight values are initialized for the consequent membership function realization unit based on knowledge held in advance.

Next, the learning process is started, the input pattern for learning is presented to the rule section prewire neuron, and the learning ratio cover pattern ( For example, weight learning using the pack propagation method is performed by modifying the initially set weight values so that they almost match the teacher signal). Do it in order.

In this way, after learning the weights of either or both of the antecedent part membership function realization part and the rule part in the rule part prewire neuron, the weights of the consequent part membership function realization part are learned. It is possible to adjust the weights of the consequent membership function realization unit in accordance with the learning results of the weights of either or both of the subject membership function realization unit and the rule unit.

In addition, the initial setting of weights that incorporates pre-existing knowledge facilitates learning and improves learning efficiency.

Furthermore, by initializing the weights using random numbers, it is possible to maximize the neural learning function.

[Embodiment] FIGS. 2A to 2E are learning operation flowcharts showing the processing procedures of the first to fifth embodiments of the learning method of the rule section pre-wire neuro according to the present invention. The rule section pre-wired neuron that is the object of the learning method according to the procedures of the first to fifth embodiments includes, for example, the antecedent section membership function realizing section 10, the rule section 12, and the consequent section membership function shown in FIG. It has a hierarchical network structure for realizing fuzzy control rules, which is composed of a realization unit 14 (including a centroid calculation realization unit 18).

Here, the rule section pre-wire neuro means the rule section 12.
has a network structure in which all units are not internally connected with the antecedent membership function realizing unit 10 in the previous stage and/or the consequent membership function realizing unit 14 in the subsequent stage, but some are internally connected according to the control rule. , is compared to a rule-part fully connected neuron that internally connects everything between hierarchical units.

In each learning process of the first to fifth embodiments shown in FIGS. 2A to 2E, the following steps 81 to S6 are performed.

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

Initialize Cbc to implement the given membership function. Note that Cab and Cbc indicate connection group numbers of connections for which weights are set, as will be described later.

[Step S2] Initialize the weights Ccd and Cde of the rule section 12 so as to realize the given fuzzy rule.

[Step S3] The weight Cef of the consequent membership function realization unit 14 is initialized so as to realize the given membership function.

[Step S4: Initialize the weight C1g of the center of gravity calculation implementation unit 18 so as to realize center of gravity calculation.

[Step S5] A phase/group correspondence table indicating neuron groups for each phase in which learning processing is performed and a group/connection correspondence table indicating connection groups belonging to each two-neelon group are set.

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

In such a learning process consisting of steps S1 to S6, the initial setting of weight values for the antecedent membership function realization unit 10, rule unit 12, and consequent membership function realization unit 14 in steps 81 to S3 is performed in each implementation. Although each example is different, in any of the examples, in step S6, first,
That is, for the part where weight learning is performed in Phase 1, initial setting of weight values based on the introduction of knowledge that is previously held or initial setting of random weight values using random numbers is performed.
For the part where the weight values are learned from the second time onwards, that is, the part where the weight values are learned from the phase 2 onwards, only the initial setting of the weight values is performed based on the introduction of previously held knowledge.

The initial setting of weight values for the center of gravity calculation implementation unit 18 in step S4 is the same for all embodiments. Further, in setting the phase/group correspondence table in step S5, a unique neuron group is set for each embodiment, but the group/connection correspondence table is set in common. Furthermore, in the weight learning by activating the learning scheduler in step S6, all embodiments have the same point in that the final stage of weight learning is performed on the consequent membership function realization unit 14. The object of weight learning performed in the preliminary stage prior to weight learning by the section membership function realizing section 14 is different for each embodiment.

Hereinafter, a specific example of the learning method of the rule section pre-wired neuron of the present invention will be explained in detail by taking the first embodiment shown in FIG. 2A as an example.

[First Example] In the first example shown in FIG. 2A, steps 81 to S3
In step S4, the weight values of the antecedent membership function realization unit 10, the rule unit 12, and the consequent membership function realization unit 14 are initialized to realize the given membership function, and in step S4, the center of gravity is calculated. After initializing the weight values of the realization unit 18, the phase/group correspondence table and the group/connection correspondence table are set in step S5, and then the weight learning of the antecedent membership function realization unit 10 is performed in phase 1 in step S6. The method is characterized in that the weight learning of the consequent membership function realization unit 14 is performed in phase 2.

Since it is the antecedent membership function realization unit 10 that first performs weight learning in phase 1 of step S6, it is necessary to initialize the weight values of the antecedent membership function realization unit in step S1. , the weight values are initialized based on the introduction of prior knowledge, or the weight values are initialized randomly using random numbers. On the other hand, for the rule section 12 that does not perform learning, the weight values are initialized based on the introduction of pre-existing knowledge as shown in step S2, and learning is performed in the final stage of phase 2. As for the consequent membership function realization section 14, as shown in step S3, similarly to the rule section 12, weight values are initially set based on the introduction of knowledge held in advance.

A detailed explanation of each process of steps S1 to S6 shown in the learning operation flow of the first embodiment is as follows.

[Configuration of the rule section prewire neuron] The rule section prewire neuron to which the present invention is applied has a hierarchical network structure capable of realizing the fuzzy control rule shown in FIG. 3, for example.

In FIG. 3, each node shown in each part represents one basic unit, and two types of basic units are used: a sigmoid function unit shown by a black circle, and a linear function unit shown by a white circle.

FIG. 4 shows the configuration of the basic unit shown in FIG. 3.

In FIG. 4, a basic unit 201 constitutes an L multi-input 1-output system, and includes a multiplication section 22 that multiplies multiple inputs by the weight values of each internal connection, and a multiplication section 22 that adds all 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, the multiplier 22
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,,=Σ'J jhw l h
(1) yp+=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 Wlh: weight value x of the internal connection between the h-i layers, I: Sum of products of inputs from each unit of the h layer to the 1st unit of the i layer Here, the above equation (2) is known as a sigmoid function, and is the threshold value processing of the basic unit indicated by the black circle shown in FIG.

In contrast, the threshold processing unit 26 ypl=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 realization section 10, does not necessarily have to be a linear function unit. The hardware of the basic unit shown in Fig. 4 is disclosed in Japanese Patent Application No. 63-216865.
"Network Configuration Data Processing Device".

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 executes each input variable Xn, X2 . ...Xn is input, input variable Xn
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 XI &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 LH3-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 calculation realization unit 16) calculates the membership function y (
Input the grade values of SA) to 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 used as 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, 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 includes four sigmoid function units 30-1 to 30.
-4 and the same four linear function units 32-1 to 32
-3.

The rule section 12 includes, for example, sigmoid function units 34-1 to 34-8 that realize eight fuzzy rules, and the antecedent membership function realization section 10 performs wiring according to the description of each fuzzy rule L1 to L8. It is. The outputs of the sigmoid function units 34-1 to 34-8 are connected according to fuzzy rules to two linear function units 36-1 and 36-2 which perform consequent operations according to fuzzy rules.

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

Furthermore, the center of gravity calculation implementation unit 18 is composed of a center of gravity determining element output device 40 and a center of gravity calculation device 42. Linear function unit of center of gravity determining element output device 40) 40-1. .

40-2 is a consequent membership function realization unit] The weighting of the unit output from 4 calculates the sum to obtain two centroid calculation elements Y, , Y2, and finally the centroid calculation device 42 calculates the two The center of gravity is calculated from the center of gravity calculation elements Y, , Y2 to generate a control output Y as a fuzzy inference value.

FIG. 5B shows neuron groups and connection groups for realizing the learning method according to the present invention for 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
As for the first one, it is excluded from the weight learning target by the backpropagation method because it merely performs separation and addition/synthesis. Therefore, there are five neuron groups that are subject to weight learning by the pack propagation method: Gb, Gc, Gd, Ge, and Gf.

On the other hand, there are six connection groups representing the input connections of each two-neelon group, Cab to Cjg, and the weight values of the input connections located before the neuron group to be learned are learned by the pack propagation method.

Regarding the learning operation flow shown in FIG. 2, which targets the rule section pre-wire neuron having the structure shown in FIGS.

[Initial setting of the antecedent membership function realizing unit] The initial setting of the weight value for the antecedent membership function realizing unit 10 is to set the weight of each connection included in the connection groups Cab and Cbc to the given membership function. Initial settings are made to achieve this.

FIG. 6 shows the antecedent membership function realization unit 10 of FIG. 5A.
The control human power x1 side in is extracted and shown. In FIG. 6, the output value y output from 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, as the weight value W of the input connection of the sigmoid function unit 30-1, W= −20 is set, and θ=−10 is set as the threshold value θ in the sigmoid function unit 3o-1.

Further, the weight value W of the input connection of the other sigmoid function unit 30-2 is set to W=20, and at the same time, the threshold value θ of the sigmoid function unit 3.2 is set to θ=10. 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 Xn(SS) for the control output Xn shown in the characteristic curve 44 in FIG. It is possible to set a truth value calculation function for a membership function having a function shape such that SS) becomes small.

Also, in the sigmoid function unit 30-2, W>Q
By initially setting the weight value and threshold value such that θ>0, the unit output value X+(LA) for the control human power Xn shown in the characteristic curve 46 in FIG. 7, that is, the larger the control human power X, the unit output value Xn(LA). It is possible to realize a function that calculates the truth value of a membership function whose function shape is such that .

Sigmoid function unit 301.30 shown in FIG.
The initial settings of the weight value and threshold for -2 are set in advance so that the desired antecedent membership function is realized for the remaining sigmoid function unit 30-3.304 side in the antecedent membership function realization unit 10. initial settings based on your knowledge.

[Initial Setting of Rule Section] The initial setting of the weight value for the rule section 12 in FIG. 5A is as follows.
The weight values of the respective connections belonging to the connection groups Cc+I and Cde shown in FIG. 5B are 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 the initial value constant of the weight value for the rule section 12 together with fuzzy rules.

In FIG. 8, eight sigmoid function units 34-
1 to 34-8 correspond to fuzzy rules L1 to L8 extracted and shown on the right side, respectively. For each of the sigmoid function units 34-1 to 34-8, fuzzy rules are calculated from the linear function units 32-1 to 32-4 of the antecedent membership function realization section 10 located at the previous stage.
Input connections are made in accordance with ~L8. Furthermore, fuzzy rules are 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 parts of 1 to L8 (T HE
Connections are made in accordance with N).

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 in the rule section 120, in which truth values produces y.

The output value y output from the sigmoid function unit 34 in FIG. 9 is as follows: 5=Wx1・Xn+Wx1・X2−θ (4) y=
It is calculated as 1/(1+e-s) (5). The unit output value y given by the above equation (5) can have an appropriate function shape by arbitrarily setting the input connection weight values Wx1 to Wx7 and the threshold value θ.

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

For example, in the sigmoid function unit 34-1 that realizes the fuzzy rule L1, the following is set, and by setting these initial values, the sigmoid function unit 34-1 that realizes the fuzzy rule L1
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 initializing the weight values and threshold values for each of the 8 fuzzy rules 2 to L8, 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 processing is a linear function, and 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 as follows: The weight values of the connections belonging to the connection group Cel in FIG. 5B are preset. 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, the linear function unit of the previous rule part 12)
Connection input is performed from 36-1 and 36-2 as shown in the figure. These nine units are 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 of the input connections for the linear function units 38-1 to 38-9 is performed by the unit 3 of the rule section 12.
Regarding the weight value of the connection of the output y (SS) from unit 6-1, follow the characteristic curve 46 of the graph shown on the right side with the horizontal axis as the control output Y1 and the vertical axis as the unit output value.
Weight values from the side to the unit 38-9 side, 1, 1. 1. G, 75. Q, 5. 0.2
5. 0. CI, initialize to 0.

Further, regarding the weight values of the input connections from the unit output y (LA) of the rule section 12, the weight values are changed from 0 to 0 from the unit 38-1 side to the unit 38-9 side according to the characteristic curve 48. 0. G, 25. Q, 5. G, '75
．． 1. ],, ■Initialize.

Furthermore, the threshold value θ of the linear function units 38-1 to 38-9 is θ
= 0.

Due to the initial setting of the weight values in the consequent membership function realization unit 14, the unit output values y(SS) and y(LA ) is reduced by multiplication by the initial weight value, added, 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 in FIG. 5A are initialized so as to realize center of gravity calculation for the connection group Cfg 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, as shown in the figure, weight values for calculating the two gravity center derived values y, , Y2 used in the gravity center calculation are set for this input connection.

In fuzzy control, a method is generally adopted in which the control output Y, which is a fuzzy inference value, is calculated by finding the center of gravity of a figure consisting of the sum of functions of reduced membership functions for the same control output position according to the following formula: do.

Y= f grade (y) ydy / f gr3d
e(y)dy (6) In order to realize this formula (6), the linear function units 40-1 are obtained from the linear function units 38 to 1 to 38-9 of the consequent membership function realization unit 14.
As the connection weight value for the connection, the minimum value Y of the control output Y =
Starting from 0, weight values are assigned that increase from 0 at equal intervals up to the maximum value Y=1. On the other hand, the linear function unit 40
-2 input connection, maximum value of control output Y = 1
From 0 to the minimum value Y=0, weight values that decrease from 0 to -1, for example, are assigned at the same equal intervals as above.

By weighting the input connections to the linear function unit 40-1, 40-2, if the unit output values of the preceding stage linear function units 38-1 to 38-9 are 01 to C9, the linear function unit 40-1 becomes , Yl=IC+ +〇, 875 C2+0.75C3〇, 625 C4+0.5 C5+[1,375C6〇
．． 25C, +0.125 C8+OC9 (7), and the linear function unit 40-2 outputs Yl”OC+
0.125 C20, 25C3-0, 375C4-0,
5C5-0,625C6-0,75C7-0,875C
Outputs 8+IC3 (8).

The centroid derived value Y, , Y2 output from the linear function unit) 40-1.40-2 is given to the centroid calculation device 42, and the centroid is calculated using That is, the control output Y, which is a fuzzy inference value, is calculated.

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

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

[Learning Schedule Setting] In step S5 of FIG. 2A, following the completion of the initial settings, a learning plan, ie, a learning schedule, for causing the rule section prewire neuro to perform learning processing is set. The setting of this learning schedule is shown in Figure 20(a).
The settings of the phase group correspondence table shown in the figure (f)
Perform the two steps shown in the group/connection correspondence table shown below. Furthermore, the second
The phase/group correspondence table in Figure 0 (b) to (e) is shown in the second
The second to fifth embodiments shown in FIGS. B to 2E are shown.

First, the phase-group correspondence table specifies the neuron groups to be learned in each phase as the learning phase progresses. In the first embodiment of the present invention, the weights of the antecedent member member function realizing unit 10 are learned after the initial setting of the weight values, and then the weights of the antecedent member member function realizing unit 14 are learned.
As shown in FIG. 20(a), in Phase 1, as shown in FIG. In addition, in Fa-22, the consequent member and the function realization unit 1 are set as the learning target group.
Neurology group Gf in FIG. 5B belonging to No. 4 is set.

On the other hand, the group connection correspondence table is set as shown in FIG. - A common connection table is set in all embodiments.

When the setting of the phase/group correspondence table and the group/connection correspondence table is completed, step S6 in FIG. 2A is completed.
Step 2 starts the learning scheduler, actually provides learning data to the rule section pre-wired neuron, and executes weight learning using the pack propagation method.

[Processing and Configuration of Learning Scheduler] FIG. 21 shows a flowchart of the learning process of the rule section pre-wired neuron of the present invention. This processing flow diagram is, for example, the 22nd
First, the device configuration for learning processing that can be realized by the device configuration shown in the figure will be explained.
In the figure, a learning instruction unit 50 and a learning unit 70 are provided for the rule section pre-wired 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, a group/connection correspondence table 58 storing a group/connection correspondence table, and a learning convergence determination section 60 are provided.

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 100, it is possible to determine whether or not the learning scheduler 52 of the learning instruction unit 50 changes the weight value of the connection for the rule section pre-wire neuro 100. It is possible to set a learning flag that determines the learning flag, and it is possible to perform weight 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 structure consisting of a function realization section, a rule section, and a consequent membership function realization section.

When a hierarchical network conforming to fuzzy rules is configured, the weight values are adjusted as follows when actually learning is performed by the pack propagation method.

In FIG. 23, the weight change information reading unit 84 of the learning adjuster 80 monitors whether the weight change amount has been written by the weight calculation unit 88 to the weight change amount storage unit 92 of the weight change unit 76. There is. 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. That is, writing to the weight change amount storage section 92 is made valid. On the other hand, if the learning flag is off, the weight change amount adjustment section 86 changes the weight change amount storage section 9 of the weight change section 76.
2 is set to zero to disable the weight change. The hardware of this learning scheduler is disclosed in Japanese Patent Application No. 63-227825.
``Learning method for network configuration data processing device'' in the issue.

[Processing contents of learning scheduler] Next, the learning process of the rule section pre-wired neuro in the first embodiment of the present invention according to the device configuration of FIGS. 22 and 23 will be explained with reference to the process flow diagram of FIG. 21. .

First, when the learning scheduler 52 of the learning instruction unit 50 is activated, step SL (hereinafter "step" 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 phase/group correspondence table 56 is referred to and the neuron group numbers Gb and Gc corresponding to the learning phase i=1 are read out. That is, the antecedent membership function realization unit 1 that performs weight learning first
The neuron group numbers Gb and Gc belonging to 0 are read out.

Next, proceed to S3, and set neuron group numbers Gb and Gc.
The connection group numbers Cab and Cbc corresponding to ``Cab'' and ``Cbc'' are read out with reference to the group/connection correspondence table 58. Next, the process advances to 84 and the connection group number Cab.

Output Cbc and connect group number Cah, Cbe
The learning flag of the learning adjuster 80 provided in the connection belonging to the connection is turned on. Specifically, the learning scheduler 52 of the learning instruction unit 50 outputs the connection group numbers Cab and Cbc to the learning instruction reading section 72 of the learning unit 70 for reading. The learning flag setting unit 74 receives the reading result from the learning instruction reading unit 72, and sets the connection group number Ca.
b. The learning flag is set by instructing the flag storage section 82 of the learning adjuster 80 provided in the connection belonging to Cbc to turn on the flag. This designates the weight learning of the antecedent membership function realizing section 10.

Next, the process proceeds to 85, where the learning scheduler 52 issues a learning execution command to start the learning process. In this learning process, the control human power X prepared in the learning signal storage section 54 and the teacher signal d, which is a desirable control output for the control human power X, are read out, and the control input is given to the rule section prewire neuro 100, and the teacher signal d is It is given to the weight changing unit 76 and the learning convergence determining unit 60. Then, the control output Y output from the rule unit prewire neuron 100 that has received the learning input 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. After making the change, when the error between the control output Y and the teacher signal d becomes less than or equal to a specified value in the learning convergence determining section 60, it is determined that learning has ended, and the learning process of the first phase is ended.

When the learning process of phase 1 is completed based on the learning execution command in step S5, the phase counter i is incremented by one in 86, and after checking the end of the learning phase in S7, the process returns to S2 again, and after corresponding to learning phase 2. The neuron group number Gf of the subject membership function realizing unit 14 is read out from the phase/group correspondence table 56.

Next, in S3, the connection group number Cel corresponding to the neuron group Gf is obtained from the group/connection correspondence table 58.
is read out, and in S4, the connection group number Cef is output, and a learning flag is set in the learning adjuster 80 provided on the connection of the consequent part membership function realization unit 14. Subsequently, in S5, a learning execution command is issued to execute weight learning of the consequent membership function realization unit 14 by the pack propagation method using the learning signal.

When the weight learning of the consequent membership function realization unit 14 in Phase 2 is completed, the phase counter i is set to 1 in S6.
Then, in S7, the end of the learning phase is determined, and the series of processing ends.

[Principle of learning by pack propagation method 1 Next, the principle of weight 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.

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. Each unit has the configuration shown in FIG. 4, and the calculations performed in the unit are as shown in equations (1) and (2) above for the basic unit 20. 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. As is clear from the above equations (1,) and (2), it is necessary to adjust the weight value W and the threshold value θi at the same time, but this work is a difficult work that interferes with each other. Therefore, in addition to the normal unit, the applicant has provided a unit whose output value is always 1 in the h layer on the input side, and uses a threshold value θi as a weight value for the input connection from this unit with an output value of 1. A pack propagation method has been proposed in which the threshold value θi can be treated as a weight value by incorporating it into the weight value Wih by allocating it (Japanese Patent Application No. 62-333484).

By doing this, the above equations (1) and (2) become y, +=
1/ (1+exp(-Xp+)) (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.

)(,,=photo VDIW+1 (
I3) yp+=1/ (1+exp(xp+))
(14) However, j is the unit number W of the 21st layer, , : the weight value x of the internal connection between the i-j layers, , : the product sum 3'p+ of the input from each unit of the i layer to the 3rd unit of the j layer '3 of the j layer for the input signal of the pth pattern
The weight changing section 76 in FIG. 22 outputs the input cover turn X for learning.
When receiving the teacher pattern apl, which is the signal to be taken by the output cover turn y, , which is output from the output layer j when δp+=]/p+ (1yp+) (apr yp
+) Calculate (15). Subsequently, ε: Learning constant ζ: Update amount ΔW of the weight value between the i layer and the j layer according to the momentum t and the number of learning times.

Calculate (1). 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 uses Δδ□ calculated by the above equation (15) to calculate δ,,=y,, (1-yp+)Σδ,, w I l
(+-1) (17), and then calculate the update amount Δwth(+) of the weight values of the h layer and the i layer.

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). To update the above weight values, output pattern y+++ or teacher pattern d for input pattern pattern for learning.
The learning process is repeated so that a weight value matching pi is obtained.

On the other hand, if the units constituting the hierarchical network of FIG. 24 are linear function units, the above equation (14) can be expressed as y=Xp+ (20). The above (15
) and (17) are δp+= d pl y pl
(21) δ0. = photo δ,, w,, (t 1 )
It appears as (22).

[Weight Learning Processing of First Embodiment] FIG. 25 is a processing flow diagram of weight learning using the pack propagation method for the pre-wired neuron of the present invention 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 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. In the first embodiment of the present invention, in need 1, antecedent membership function realization unit 1
0 becomes the weight learning target layer, and in the next phase 2, the consequent part membership function realization unit 14 becomes the weight learning target, and the weights of the other rule unit 12 and centroid calculation realization unit 18 are excluded from the learning target. has been done.

Here, the fuzzy inference value output by the realization of the fuzzy rule by the pre-wired neuron in the rule section, that is, the control output value, is y6, and the learning process in one phase is performed from the 6th layer to the 5th layer, the 4th layer, and the 4th layer. The third layer, the second layer, and the first layer are performed in this order.

Of course, if the fifth layer belonging to the consequent membership function realization unit 14 is the learning target layer as in phase 2, the process ends with the fifth layer.

The progress of learning for each layer in one phase is indicated by an i counter. That is, in the initial state, the i counter is set to 1=6.

To further simplify the explanation, Wl, l-1 and ΔW represent a matrix, the size of which is (number of neurons in the first layer) x (number of neurons in the i-1st layer). Furthermore, δ1 and yi represent vectors whose size is the same as the number of neurons in the i-th layer.

The weight learning flow shown in FIG. 25 for the first embodiment of the present invention will be explained as follows based on the parameters of each part shown in FIG. 26.

First, a counter for setting a learning target layer is initialized to 1 using Sl. Specifically, since the device has a six-layer structure as shown in FIG. 26, the counter i is initially set to i=6. With this initial setting, the unit in the sixth layer is designated as a learning target.

Next, the process proceeds to step 32, where it is determined whether the unit is a sigmoid function unit. In this case, since it is a linear function unit, 814
The process proceeds to step 315, where it is determined whether the layer is the final layer or not. Since the sixth layer is the final layer, the process proceeds to step 315, and the above equation (21) is calculated using the teacher signal d and control output y6 obtained at that time. The difference value δ6 is determined according to a formula similar to the above.

Next, the process proceeds to 86, where it is determined whether the unit is to learn weights. Since the sixth layer is excluded from the learning target, specifically, the learning flag of the learning adjuster 80 provided in the connection is turned off, so the process proceeds to S8 and the weight value update amount ΔW6
The process proceeds to S9 as 5-0, and the update of the weight value W65 is invalidated. Next, the process advances to S12, where counter j is set to phase 1.
It is determined whether the learning end layer E = 1 or not, and since i = 6 at this time, the process proceeds to 313, decrements the counter i by 1, sets i = 5, and returns to S2 to start learning the next 5th layer. enter.

Since the fifth layer is also a linear function unit, 82
Proceeds to 814, but since it is not the final level, it goes to 31
Proceed to step 6, and calculate the difference value δ according to a formula similar to formula (22) above.
, is calculated. Next, the process proceeds to 86 to determine whether or not it is a learning target, and since the fifth layer is not a learning target in phase 1, the process proceeds to S8 to update the weight value update amount ΔW 54
= O to invalidate the weight update in S9. Similarly, the same process is repeated until the final unit of the fifth layer, and when the process of the fifth layer is completed, the process proceeds to the fourth layer. Since the fourth layer is also a linear function unit and is excluded from the phase 1 learning target, the same process as the fifth layer is repeated.

Further, processing of the third layer proceeds from 82 to S3 since the third layer uses a sigmoid function unit, and since it is not the final layer unit, the difference value δ3 unique to the sigmoid function unit is calculated from the above ( 17) Calculate using the same arithmetic expression as Equation 17).

This third layer is also not a learning target in Phase 1, so after setting the weight value update amount ΔW to 320 in S6 to 88,
The weight update in S9 is invalidated. Thereafter, the same process is repeated until the final unit of the third layer, and the process proceeds to the second layer.

Since the second layer is a sigmoid function unit and is the learning target in Phase 1, the difference value δ2 is calculated in S16 via 82° S4 according to an equation similar to equation (22) above.
After calculating the weight value update amount ΔW, the process proceeds from S6 to 87.
21 is calculated according to equations similar to equations (16) and (18) above, and in S9, the weight value ΔW21 is updated according to an equation similar to equation (19) above. Similarly, the same process is repeated until the final unit of the second layer, and when the process of the second layer is completed, the process proceeds to the process of the first layer.

The first layer is a sigmoid function unit, and since it is the learning target of phase 1, it passes from 82 to 83 to S5
The process proceeds to calculate the difference value δ1 specific to the sigmoid function unit, the process proceeds from S6 to 87 to obtain the weight value update amount ΔW10, and the weight value update amount ΔW, in S9. The weight value W1 is obtained by adding
Update 0. Thereafter, the same processing is repeated until the final unit of the first layer, and when the processing of the final unit of the first layer is completed and the process proceeds from SIO to 812, at this time the counter i is i=
1 and the end layer E=1, it is determined that all learning has been completed, and the process of phase 1 is ended.

Next, the weight learning process of phase 2 will be explained. In phase 2 of the first embodiment, the fifth layer belonging to the consequent membership function realization unit 14 is the learning target. Therefore, the learning end layer E at 312 in FIG. 25 is set to E=5.

When weight learning in Phase 2 is activated, the learning process starts from the 6th layer, as in Phase 1, and when the 6th layer is completed and the process proceeds to the 5th layer, the 5th layer becomes the learning target. Since it is a linear function unit, the weight value update amount ΔW,4 is calculated in S7 based on the difference value δ obtained in 816, and the weight value W,4 is updated in S9. Similarly below,
If you process up to the final unit of the 5th layer, it will be 8 from SIO.
The process proceeds to step 12, and since E=5 is set at this time, when the counter i=5 matches, it is determined that the learning of all the layers in phase 2 has been completed, and the series of processing ends.

Next, second to fifth embodiments of the present invention will be described.

[Learning Process of Second Embodiment] In the second embodiment, as shown in step S6 in FIG. 2B, weights are learned in the following order: (1) rule section 12 (2) consequent section membership function realization section 14. Corresponding to this weight learning, the rule unit 12 that first learns the weights initializes the weight values based on the introduction of pre-existing knowledge, or randomly sets the weight values using random numbers, as shown in step S2. The weight values are initialized, or both are used together to initialize the weight values. On the other hand, for the antecedent membership function realization unit 10 that does not learn weights and the consequent membership function 14 that learns weights the second time, as shown in steps SL and S3, pre-existing knowledge is introduced. The weight values are initially set based on

The setting of the phase group correspondence table in step S5 is
As shown in FIG. 0 (b), the neuron group number Gd belonging to the rule section 12 in phase 1.

Ge is set, and in phase 2, the neuron group number Gf belonging to the consequent membership function realization unit 14 is set. Incidentally, the group/connection correspondence table is as shown in FIG. 20(f), similar to the first embodiment.

The details of the learning process of the second embodiment can be immediately inferred from the explanation of the first embodiment described above.

According to this second embodiment, by learning the weights of the consequent membership function realizing unit 14 after learning the weights of the rule unit 12, the consequent membership function is adjusted to match the result of weight learning of the rule unit 12. It becomes possible to adjust the weights of the function realization unit 14.

[Learning process of third embodiment] In the third embodiment of the present invention, step S in FIG. 2C
As shown in 6, weight learning is performed in the following order: (1) rule unit 12 (2) antecedent membership function realization unit 10 (2) consequent membership function realization unit 14. Corresponding to the order of weight learning, the rule section 12 that learns the weights first sets the weights to initial values based on the introduction of pre-existing knowledge, or uses random numbers, as shown in step S2. The weight values are initially set randomly using the method, or the weight values are initially set using a combination of both methods. On the other hand, for the antecedent membership function realization unit 10 and the consequent membership function realization unit 14 that learn weights from the second time onward, step S1;
As shown in S3, weight values are initialized based on the introduction of knowledge held in advance.

The phase/group correspondence table set in step S5 is as shown in FIG. 20(C).
Neuron group numbers Gd and Ge belonging to the antecedent membership function realization unit 10 are set in phase 2, and neuron group numbers Gb and Gc belonging to the antecedent membership function realization unit 10 are set in phase 3. The neuron group number Gf to which it belongs is set.

According to this third embodiment, after learning the weights of the rule unit 12 and the antecedent membership function realization unit 10, the weights of the consequent membership function realization unit 14 are learned, so that
The weights of the consequent membership function realization unit 14 are adjusted in accordance with the weight learning results of the antecedent membership function realization unit 10 and the rule unit 12.

[Learning process of the fourth embodiment] In the fourth embodiment of the present invention, step S in FIG.
As shown in 6, the weights are learned in the following order: (1) Antecedent membership function realization unit 10 (2) Rule unit 12 (2) Consequent membership function realization unit 14. For the antecedent membership function realizing unit 10 that learns the weights first in accordance with this weight learning order, the weight values are initialized based on the introduction of pre-existing knowledge as shown in step S1, or Alternatively, the weight values are initialized randomly using random numbers, or the weight values are initialized using a combination of both. Further, regarding the rule section 12 and the consequent membership function realization section 14 that learn the weights from the second time onward, step S2. As shown in S3, the weight values are initialized based on the introduction of knowledge held in advance. Furthermore, the setting of the phase group correspondence table in step S5 is as shown in FIG. 20(d).
and neuron group Gb belonging to the antecedent membership function realization unit 10.

Gc is set, neuron groups Cyd and Ge belonging to the rule section 12 are set in phase 2, and further phase 3
The neuron group Gf belonging to the consequent membership function realization unit 14 is set.

According to the learning process of the fourth embodiment, by learning the weights of the consequent membership function realizing unit 14 after learning the weights of the antecedent membership function realizing unit 10 and the rule unit 12, The weights of the consequent membership function realization section 14 can be adjusted by combining the weight learning results of both the subject membership function realization section 10 and the rule section 12.

[Learning Process of Fifth Embodiment] In the fifth embodiment of the present invention, as shown in step S6 in FIG. The weights are learned in the order of 14. Corresponding to the order of weight learning, the antecedent membership function realizing unit 10 and the rule unit 12, which learn weights first, calculate weight values based on the introduction of pre-existing knowledge as shown in step 31.82. The weight values may be initialized, or the weight values may be initialized randomly using random numbers, or a combination of both may be used to initialize the weight values. - On the other hand, for the consequent membership function realization unit 14 that performs the second weight learning, weight values are set based on the introduction of knowledge held in advance, as shown in step S3. Furthermore, as shown in FIG. 20(e), the setting of the phase-group correspondence table in step S5 is such that in phase 1, neuron groups Gb and Gc belonging to the antecedent membership function realizing section 10 and neurons belonging to the rule section 12 are set. Groups Gd and Ge are set, and in phase 2, a neuron group Gf belonging to the consequent membership function realization unit 14 is set.

[Others] The above specific embodiment is as shown in FIG. 5A.
Although the pre-wired neuron rule unit that realizes fuzzy rules targeting two control inputs Xn and X2 has been taken as an example, it goes without saying that the number of control inputs may be 1 or 3 or more as appropriate. . In addition, in the above embodiment, y=1/ (1+e-"") (23) shown in FIG.
As shown in FIG. 27, two sigmoid function units 94 are used as an example.
The unit output of -1.94-2 may be given to the subtraction unit 96 to realize the function shape shown in FIG. In this case, the unit output y is y=1/(1-e-”””)-1/(1-e-w2
”#2).

[Effects of the Invention] As described above, according to the present invention, after learning the weights of either or both of the antecedent membership function realization part and the rule part, the weights of the consequent membership function realization part are learned. By learning the following, it is possible to achieve optimal weight adjustment of the consequent membership function realization unit in accordance with the weight learning result of either or both of the antecedent membership function realization unit and the rule unit.

In addition, weight learning can be easily performed by initializing the weight values based on the introduction of pre-existing knowledge.

Further, by initializing the weight values using random numbers or in combination with pre-existing knowledge, it is possible to perform learning that maximizes the functions of the neuron network.

4. II! Brief explanation of J side Figures 1A, IB, IC, ID, and IE are explanatory diagrams of the principle of the present invention; Figures 2A, 2B, 2C, 2D, and 2E are learning operation flow diagrams of the present invention; Figure 3 is the book FIG. 4 is an explanatory diagram of the basic unit; FIG. 5A is a configuration diagram of a specific example of the rule section pre-wire neuron of the present invention; FIG. 5B is FIG. 5A Fig. 6 is an explanatory diagram of the initial settings of the antecedent membership function realization section of the present invention; Fig. 7 is an explanatory diagram of the membership function of the antecedent part; Fig. 9 is an explanatory diagram of the rule section unit of the invention; Fig. 10 is an explanatory diagram of fuzzy rule upper 1 in Fig. 9; Fig. 11 is an illustration of fuzzy rule upper 2 in Fig. 9.
Explanatory diagram Figure 12 is an explanatory diagram of the top 3 fuzzy rules in Figure 9;
Figure 13 is an explanatory diagram of the upper four fuzzy rules in Figure 9;
Figure 15 is an explanatory diagram of Fuzzy Rule Upper 5 of Figure 9; Figure 16 is an explanatory diagram of Fuzzy Rule Upper 6 of Figure 9; Figure 17 is an explanatory diagram of Fuzzy Rule Upper 6 of Figure 9; An explanatory diagram of fuzzy rule L8; FIG. 18 is an explanatory diagram of the initial settings of the consequent part membership function implementation unit of the present invention; FIG. 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 learning process flow diagram of the present invention; Fig. 22 is a diagram illustrating the device configuration of the learning process according to the present invention; Fig. 23 is Fig. 22 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 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 part 16: input part 18: centroid calculation realization part 20: basic unit 22: multiplication part 24: addition part 26: threshold processing part Function realization unit 28-L 2g-2, 32-1 to 32-4.38-1
~3g-940-140-2+linear function unit 30-1~3Q-2, 34-1~34-1194-19
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 Duplex change 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 Two subtraction unit 100: Rule section pre-wire neuron

Claims (7)

[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 units with 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 a learning method of a rule section pre-wired neuro that is internally coupled and 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) initializes weight values based on knowledge held in advance or using random numbers, and the rule unit (12) and the consequent membership function realization unit (14) ); a second process of learning the weights of the antecedent membership function realizing unit (10) using learning data; and learning. The consequent membership function realization unit (14) using data
A learning method for a rule section pre-wire neuro, comprising: a third step of learning the weights of; (2) It has a hierarchical network consisting of an antecedent membership function realization unit (10), a rule unit (12) having one or more layers, and a consequent membership function realization unit (14), and the rule unit (12) does not internally connect all units with 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 a learning method of a rule section pre-wired neuro that is internally coupled and outputs one or more control operation quantities (Y) corresponding to input control state quantities (X_1, X_2,...X_n), The rule section (12) initializes the weight values based on knowledge held in advance or using random numbers, and the antecedent membership function realization section (10) and the consequent membership function realization section (14) ); a second step of learning the weights of the rule section (12) using the learning data; and a second step of learning the weights of the rule section (12) using the learning data; A method for learning a rule part prewire neuro, comprising: a third step of learning the weights of a consequent part membership function realizing part (14); (3) 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 units with 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 a learning method of a rule section pre-wired neuro that is internally coupled and outputs one or more control operation quantities (Y) corresponding to input control state quantities (X_1, X_2,...X_n), The rule section (12) initializes the weight values based on knowledge held in advance or using random numbers, and the antecedent membership function realization section (10) and the consequent membership function realization section (14) ); a second step of learning the weights of the rule section (12) using the learning data; and a second step of learning the weights of the rule section (12) using the learning data; a third step of learning the weights of the antecedent membership function realization unit (10); and the consequent membership function realization unit (14) using the learning data.
A learning method for a rule section prewire neuro, comprising: a fourth step of learning weights of; (4) It has a hierarchical network consisting of an antecedent membership function realization unit (10), a rule unit (12) having one or more layers, and a consequent membership function realization unit (14), and the rule unit (12) does not internally connect all units with 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 a learning method of a rule section pre-wired neuro that is internally coupled and 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) initializes weight values based on knowledge held in advance or using random numbers, and the rule unit (12) and the consequent membership function realization unit (14) ); a second process of learning the weights of the antecedent membership function realizing unit (10) using learning data; and learning. a third step of learning the weights of the rule section (12) using data; a fourth step of learning the weights of the consequent membership function realization section (14) using the learning data; A learning method for the rule section pre-wired neuro. (5) 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 units with 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 a learning method of a rule section pre-wired neuro that is internally coupled and 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) and the rule unit (12) initially set weight values based on knowledge held in advance or using random numbers, and the consequent membership function realization unit (14) ); a first step of initializing weight values based on knowledge previously held in ); simultaneously learning the weights of the antecedent membership function realization unit (10) and the rule unit (12) using learning data; a second step of learning the weights of the consequent membership function realizing section (14) using learning data; . (6) In setting the initial weight values in the first process of learning the weights, weights that can be set based on prior knowledge are initialized using this knowledge, and other weights are initialized using random numbers. A learning method for a rule section pre-wired neuron according to any one of claims 1 to 5. (7) At the start of learning of each process performed after the initial setting of the weight values of the first process, the antecedent membership function realization unit (
10) Turn on the learning flag provided for each connection in the part that is the learning target in the rule part (12) and the consequent part membership function realization part (14), and check the connections that are not the learning target. 6. The rule section pre-wire according to claim 1, wherein a learning flag provided for each connection in the part is turned off, and only weight values of connections having learning flags in an on state are optimized through learning processing. How to learn neuro.