JPH03271803A - Fuzzy controller - Google Patents

Abstract

PURPOSE:To freely change the arithmetic contents of an antecedent part operation by allowing an antecedent part arithmetic part to constitute an internal coupling network of plural fundamental units for receiving an internal state value and obtaining a sum-of-products value, and also, obtaining an output value by executing the function conversion of the sum-of-products value. CONSTITUTION:An antecedent part arithmetic part 17 is constituted of a network structure part 20 which is constituted by internal coupling of plural fundamental units for receiving one or plural inputs and an internal state value multiplied against the input and obtaining a sum-of-products value, and also, obtaining an output value by executing the function conversion of the sum-of-products value an executes an antecedent part operation, and an internal state value managing part 21 for managing the internal state value allocated to the internal coupling of this network structure part 20. In such a way, even if a subroutine is not prepared in advance in the device, the arithmetic contents of the antecedent part arithmetic part 17 are changed, and it is possible to cope easily with a change request of an antecedent arithmetic function described in a fuzzy control rule.

Description

[Detailed Description of the Invention] C Overview] Fuzzy control rules In order to be able to change the function flexibly, the antecedent part calculation means that executes the antecedent calculation function is connected to an input unit that distributes the input signal value, one or more inputs from the previous layer, and an input unit that distributes the input signal value. A basic unit that receives an internal state value to be multiplied by human power to obtain a sum of products value, and obtains an output value by converting the sum of products into a function is used as a constituent unit, and multiple input unino F are one or more intermediate layers with an input layer and one or more basic units as an intermediate layer, and one or more basic units as an output layer, between the input layer and the foremost intermediate layer, Intermediate layer mutual distance [
Field of Industrial Application] The present invention relates to a fuzzy controller that calculates and outputs a control operation amount corresponding to a control state quantity by executing fuzzy inference according to a fuzzy control rule, and in particular, relates to a fuzzy controller that calculates and outputs a control operation amount corresponding to a control state quantity by executing fuzzy inference according to a fuzzy control rule. Fuzzy control is becoming popular as a new control processing method. This fuzzy control expresses a control algorithm that includes ambiguity such as human judgment in a 1f-then format, and executes this @ control algorithm according to fuzzy reasoning to calculate the control operation amount from the detected control state amount. The control target is controlled by calculating the In order to absorb the ambiguity of the control algorithm, the fuzzy controller that realizes this fuzzy control needs to have a configuration Cτ that allows the /ji arithmetic function described in the control algorithm to be easily changed9. [Prior art] The fuzzy control rule is ~ if y, i:, big and X2is small
l then y, is big (The iF part is called the antecedent part, and is the part that describes the conditions for controlling state quantities such as temperature data, and the THFN part is called the consequent part, This is the part that describes the conditions for the control operation amount of the operating end, etc.). In addition to managing multiple fuzzy control rules, a configuration is adopted in which the meanings of ambiguous linguistic expressions such as "large" or "small" written in fuzzy control rules are quantified and managed as functions. , When a fuzzy controller is given a control state quantity such as temperature data or water data from a controlled object, it first calculates the Menharsimp function (antecedent part Menharsimp function) of the control state quantity being managed. The membership function (truth value) of the given control state quantity is calculated from the following.

Next, in accordance with antecedent part calculations such as selecting the minimum value, a process is executed to determine a common value for the consequent part in each fuzzy control rule. That is, to explain using the example of the fuzzy control rule mentioned above, rx, isbig +
The truth value of is "0,8", and [XZ is sm
If the truth value of "all" is 0.5", then this "0.5" is applied to the consequent of the fuzzy control rule according to the antecedent operation which selects the minimum value -1. Next, the fuzzy controller processes the same member Jimbu function (consequent member member) of the same control amount according to the consequent operation such as selecting the maximum value. 2 for each fuzzy control rule's consequent B, which is a
ry of the consequent, is bigJ using the fuzzy control rule described above.
``0.5'' is the applied value for
0.6" as the applicable value, this "0.6" is the common value for the 'big' membership function of the control manipulated variable y1 according to the consequent operation that selects the maximum value. It is processed to make a decision.

Subsequently, the fuzzy controller reduces the membership function of the control manipulated variable according to the determined application value, and
In accordance with the process of finding the center of gravity of the figure of the function sum of this reduced membership function for the same control operation amount, the process of calculating the control operation amount which is a fuzzy inference value is executed and output to the operation terminal etc. This process will be executed.

That is, the membership function of r big J of the control manipulated variable yI is reduced according to its determined application value, and the membership function such as 's+l1alJ of the control manipulated variable yl obtained from another fuzzy control rule is reduced according to its determined application value. In addition to reducing the size according to the value, the control operation amount that is the output of the fuzzy controller is calculated by calculating the center of gravity of the figure of the sum of the membership functions.

Conventional fuzzy controllers have adopted a configuration in which execution of such fuzzy control rules is achieved by executing them programmatically. That is, according to the software means of the computer system that performs sequential processing, first, the truth value of the control state quantity is calculated, then the applied value for the consequent is determined according to the antecedent part operation, and then Then, the common value of the control operation amount with respect to the membership function is determined according to the consequent part calculation 2, and then the control operation amount is calculated and outputted by a center of gravity calculation or the like.

[Problem to be solved by the invention]

However, fuzzy control rules are generated according to the knowledge of the target process of an operator who is familiar with controlling the target.

The reality is that it is not possible to generate fuzzy control rules that can achieve the desired control from the beginning, and we will tune them through trial and error while evaluating the generated fuzzy control rules through simulations and on-site tests. Therefore, it is necessary to complete the fuzzy control rules that are suitable for the object to be controlled.

From now on, there may be cases where it would be more appropriate to use a different antecedent operation than the originally assumed arithmetic function.

However, if a conventional method of configuring a fuzzy controller using software means is adopted, various antecedent functions must be prepared in advance as subroutines in the device in order to meet such requirements. be. However, since only one antecedent calculation function is originally used, such a configuration poses the problem of wasteful use of memory capacity. Moreover, if a fuzzy controller that has already been shipped does not have a subroutine for its arithmetic function, a problem arises in that it is difficult to implement an appropriate fuzzy control logic.

The present invention has been made in view of the above circumstances, and it is possible to perform the antecedent calculation function described in the fuzzy control rule without increasing the memory capacity (and without causing an increase in memory capacity). The purpose is to provide a new fuzzy controller that can be changed.

[Means to solve the problem] FIG. 1 shows the principle configuration circle of the present invention.

In the figure, 10 is a fuzzy controller equipped with the present invention, 11 is a fuzzy rule management unit that manages fuzzy control rules that describe control logic for a controlled object, and 12 is a fuzzy rule selection unit. The processing target is the fuzzy control rule managed by the fuzzy rule management section 11 (13 is the antecedent membership function management section and the fuzzy control rules managed by the fuzzy rule management section 11 are the processing targets). 14 is a consequent membership function management unit that manages the truth value of the antecedent membership function that quantifies the linguistic expression regarding the control state quantity. 15 is an input data reception unit that manages the truth value of the consequent function that quantifies the linguistic expression regarding the amount of operation; 16 is an antecedent truth value calculation unit, which calculates the truth value of the antecedent Menno\-Simbu function of the input control state quantity using the selected fuzzy control rule as a processing unit. 17 is an antecedent calculation unit,
18 determines the applied value for the consequent of the fuzzy control rule to be processed by performing an antecedent computation on the truth value of the calculated antecedent member jump function; Part calculation part that determines the applied value for the consequent part membership function from the applied value for the consequent part of each fuzzy control rule given for the same consequent part membership function. , 19 is an inference value calculation unit, which reduces the consequent membership function for the same control operation amount according to the applied value determined by the consequent calculation unit 18, and calculates the reduced consequent. Control operation amount data, which is a fuzzy inference value, is calculated by finding the graphic center of gravity of the sum of the partial Menharsimp functions.

The antecedent part calculation unit 17 of the present invention receives one or more inputs and an internal state value by which the inputs are to be multiplied, obtains a product sum value, and converts the product sum value into a function. A network structure section 20 is configured by internally connecting a plurality of basic units 1 to obtain an output value, and executes an antecedent operation.
and an internal state value management section 21 that manages internal state values assigned to internal connections of this network structure section 20. This network structure section 20 has a basic unit 1 and an input unit 1' that distributes input signal values as constituent units, and a plurality of input units 1'-
an input layer composed of h, an intermediate layer composed of one or more basic units (ii) and provided in one or more stages, and an output layer composed of one basic unit 1-j. In addition, a hierarchy that mutually internally connects between the input unit 1''-h and the basic unit 1-1, between the basic units 1-1, and between the basic unit l-i and the basic unit 1-j. A network structure may be adopted.Furthermore, a value is set as the internal state value managed by the internal state value management unit 21 that causes the network structure unit 20 to operate to output the minimum value of the input signal value. Sometimes.

(Operation) The antecedent part calculation unit 17 of the present invention is configured by a network structure unit 20 that is configured of, for example, a hierarchical network structure. According to the data conversion function defined by the assigned internal state value, the process of converting the incoming input signal into the corresponding output signal is executed.From now on, the antecedent part truth value calculation unit 1
When the truth value of the antecedent membership function for the fuzzy control rule selected from 6 is given as an input signal, the network structure unit 20 outputs the corresponding output signal as the result value of the antecedent operation. It will work like this.

The data conversion function of the network structure section 20 can freely change the calculation content by changing the value of the internal state value assigned to the internal connection even if the seven node work structures are the same.

That is, when the learning value of the internal state JJ value at which the network structure unit 20 operates to output the minimum value of the input signal value is set in the internal state value management unit 21, the antecedent part calculation unit 17:
The calculation process of finding the minimum truth value of the antecedent function for the calculated fuzzy control rule is executed, and the network structure unit 20 calculates the algebra of the input signal value in the internal state nu management unit 21. When the learning value of the internal state value that operates to output the product is set, the antecedent part calculation unit 17 calculates the algebraic product of the truth values of the antecedent membership function for the calculated fuzzy control rule. As in performing calculation processing,
By simply changing the value of the internal state value registered in the internal state value management section 21, the calculation contents of the antecedent section calculation section 17 can be freely changed.

As described above, according to the present invention, the antecedent calculation section 17 can be freely executed even if it is not prepared in advance as a subroutine in the device.
Since the content of the calculation can be changed, it is possible to easily respond to a request to change the antecedent part calculation function described in the fuzzy control rule. From now on, it will be possible to improve the performance of fuzzy controllers.

〔Example〕

Hereinafter, the present invention will be described in detail according to an embodiment applied to a hierarchical network configuration data processing device.

FIG. 2 illustrates an embodiment of the present invention. In the figure, the same parts as those explained in FIG. 1 are indicated by the same symbols. Reference numeral 20a denotes a hierarchical network structure unit that executes the antecedent computation, which corresponds to the network structure unit 20 in FIG.
A weight value management unit 21a that manages weight values assigned to internal connections of 0a, which is similar to the internal state value management unit 2 in FIG.
This corresponds to 1.

As explained in FIG. 1, the antecedent truth value calculation unit 16 performs processing to calculate the truth value of the antecedent membership function of the input control state quantity. That is, the antecedent part of the selected fuzzy control rule is the control state quantity
,,X,,X,,IF L is Small
+ X2 is Mediu+a, -, Li is B
ig, and from the controlled object, X l = A + , X z = A t ,
. . , XA7, the antecedent truth value calculation unit 16 calculates a member that quantifies rSma 11 J of the control state amount X, as shown in FIG. The membership function value Z1 of the value A is specified according to the ship function, and the control state quantity X2 (7
) rMedium", and identify the membership function value Z2 that the value A2 has, and determine the membership function value Z2 that the value A2 has according to the membership function that quantifies the control state quantity Perform processing to specify Z, . In addition, in FIG. 2, for convenience of explanation, these membership function values (truth values) are shown as being found in parallel, but in reality, they are found one by one in time series.

The hierarchical network structure section 20a performs an antecedent section operation on the truth value given from the antecedent section truth value calculation section 16, and calculates the consequent section of the fuzzy control rule for which the truth value is calculated. It will be processed to determine the applicable value for.

FIG. 3 illustrates the basic configuration of the basic unit 1 that will constitute the hierarchical network structure section 20a. As shown in this figure, the basic unit 1 is a multi-car output system, and includes a multiplication processing section 2 that multiplies multiple inputs by the weight values of their respective internal connections, and a multiplication processing section 2 that adds all the multiplication results. It is equipped with an accumulation processing section 3 and a dark value processing section 4 that performs nonlinear dark value processing on the accumulated value and outputs one final output, and the h layer is the first layer and the i layer is the first layer. In the latter layer, the accumulation processing section 3 of the i-th basic unit 1 of the i-th layer executes the calculation of the following formula (1), and the dark value processing section 4 executes the calculation of the following formula (2). Process as follows.

Xpi=Σ)'phWth (1
) formula Vpt=1/(1+exp(-Xpi+θi)
) (2) where, h: Unit number of h layer p: Input signal pattern number θ,: Darkness value of the 1st unit of i layer W4b: Weight value of internal connection between h-i layer y□: p-th The output from the h-th unit of the h-layer in response to the input signal of the pattern.The hierarchical network structure section 20a also has an input for distributing the truth values calculated by the basic unit 1 having such a configuration and the antecedent truth value calculation section 16. As shown in FIG. 4, the input layer is composed of a plurality of input units 1-i, and one or more basic units 1-i. It includes an intermediate layer provided in one or more stages (one stage in this embodiment) and an output layer constituted by one basic unit 1-j, and an intermediate layer provided between the input units 1-h and the basic unit Li. , basic unit) Li, and basic unit 1-4
It is configured to have a hierarchical network structure formed by mutually internally connecting the basic units 1-j and the basic units 1-j. Here, the number of input units 1°-h is such that the antecedent part 17 is divided into one hierarchical network structure part 20a.
When implementing with 9. The fuzzy control rule that describes the largest number of antecedent conditions is prepared in accordance with the number of conditions (that is, the required number of truth values).

The hierarchical network structure unit 20a configured in this way converts an input signal into a corresponding output signal according to a data conversion function defined by a hierarchical network structure and a weight value assigned to an internal connection of the hierarchical network structure. , and the antecedent operation is executed according to this data conversion function.

The hierarchical network structure section 20a can be constructed by either a program means or a hardware means, but in the case of constructing it by a hardware means, it can be constructed as described in Japanese Patent Application No. 1983-216865 filed by the present applicant. Application No. 318, August 1988, “Network configuration data processing device”
)” can be used.

That is, as shown in FIG. 5, the basic unit 1 is a multiplication type D that multiplies the output from the previous layer input via the input frame section 7 by the weight value held by the weight value holding section 8.
/A converter 2a, an analog adder 3a that adds the output value of the multiplication type /A converter 2a and the previous cumulative value to calculate a new cumulative value, and an analog adder 3a that calculates a new cumulative value. A sample hold circuit 3b that holds the data, a nonlinear function generation circuit 4a that nonlinearly converts the data held in the sample hold circuit 3b when the accumulation process is finished, and an analog of the nonlinear function generation circuit 4a that outputs to the subsequent layer. This is realized by including an output holding section 5 that holds a signal value, an output switch section 6 that outputs the data held by the output holding section 5, and a control circuit 9 that controls each of these processing sections.

The hierarchical network structure section 20a is realized in a configuration in which basic units 1 having this configuration are electrically connected through one common analog bus 70, as shown in FIG. Here, in the figure, 71 is a weight output circuit that gives a weight value to the weight holding section 8 of the basic unit 1, 72 is an initial signal output circuit corresponding to the input unit 1, and 73 is a synchronous control door that is a control signal for data transfer. A synchronous control signal line 74 that transmits signals to the weight output circuit 71, the initial signal output circuit 72, and the control circuit 9 is a main control circuit that sends out a synchronous control signal.

In the hierarchical network structure section 20a having this configuration, the main control circuit 74 sequentially selects the basic units 1 of the previous layer in chronological order, and in synchronization with this selection process, the output holding section of the selected basic unit l. 5 is processed so as to be outputted to the multiplication type D/A converter 2a of the basic unit 1 in the subsequent layer via the analog bus 70 in accordance with the time-division transmission format. Upon receiving this input, the multiplication type D/A converter 2a of the basic unit 1 in the subsequent layer sequentially selects the corresponding weight value, multiplies the input value and the weight value, and connects the analog adder 3a and sample hold. The accumulation processing unit 3 configured with the circuit 3b sequentially accumulates the multiplication values, and when all the accumulation processing regarding the basic unit 1 of the previous layer is completed, the main control circuit 74 starts the nonlinear function generation circuit 4a of the basic unit 1 in the subsequent layer to calculate the final output, and the output holding unit 5 performs processing to hold the final output of the conversion processing result. Then, the main control circuit 74 uses this subsequent layer as a new previous layer and repeats the same process for the next subsequent layer so that an output cover turn corresponding to the input cover turn is output. .

The hierarchical network structure unit 20a is characterized in that even if the hierarchical network structure is fixed, its data conversion function can be freely changed by changing the weight values assigned to internal connections of the hierarchical network structure. There is.

FIG. 7 illustrates an embodiment of a learning system for learning weight values assigned to internal connections of this hierarchical network structure section 20a.Next, according to an embodiment of this learning system, the hierarchical network structure section 20a The learning process for determining the weight value of the inner connection will be explained in detail.

In FIG. 7, 30 is a hierarchical network simulator constructed on a computer, which simulates the data processing function of the hierarchical network structure section 20a, and 40 is a learning signal presentation device, which is a hierarchical network structure section 20a.
A learning signal group (a pair of a learning presentation signal and a learning teacher signal is used as a basic unit) used for learning the weight value of the internal connection of a is managed, and the learning presentation signal in the learning signal group is managed. The signal is presented to the hierarchical network simulator 30, and a learning teacher signal is presented to the learning processing device 50, which will be described later. In order to realize this process, the learning signal presentation device 40 includes a learning signal storage section 4 that manages a group of learning signals.
1, and a learning processing device which reads learning signals for learning from the learning signal storage unit 41, presents a learning presentation signal among them to the hierarchical network simulator 30, and generates the other learning teacher signal forming the pair, which will be described later. 5o and a learning signal presentation unit 42 that presents the learning signal to the learning convergence determination unit 43, which will be described next.
Based on the output signal outputted from the hierarchical network simulator 30 and the learning teacher signal from the learning signal presentation unit 42, it is determined whether the error of the data processing function of the hierarchical network simulator 30 is within an allowable range. and a learning convergence determination section 43 that notifies the learning signal presentation section 42 of the determination result.

Reference numeral 50 denotes a learning processing device that implements a back propagation method known as a learning algorithm for weight values of a hierarchical network configuration data processing device, and outputs the learning presentation signal in response to presentation of a learning presentation signal by the learning signal presentation device 40. The error value between the output signal group from the hierarchical network simulator 30 and the learning teacher signal group presented from the learning signal presentation device 40 is calculated, and the weight value to be studied is calculated based on the error value. In other words, the learning processing device 50 learns the weight value of the internal connection for which the error value falls within the allowable range by sequentially updating the h layer shown in FIG. 4 according to the back propagation method. To describe the hierarchical network structure unit 20a having a three-layer structure of ~i layer and j layer, the output signal y□ from the output layer that is output when a learning input signal (learning presentation signal) is presented, When the learning teacher signal d pj, which is the signal that the output signal yp should take, is determined, first, the output signal 7DJ and the teacher signal d are determined.
Calculate the difference value (dpJ y*=) with pj, and then
αm;=Vei (17pJ) (day Vpj) is calculated, and then ΔwJI(g=εΣα□y□+ζΔWJt(t 1)
Accordingly, the update amount of the weight value between i and jN is △W
Calculate (1). Here, t represents the number of times of learning, and the reason for adding the update amount of the weight value determined at the previous update cycle is to speed up the learning. is the calculated α2. First, βpt=ypt(1-yst)Σα,,W,1(
t-1) in summer, then ΔWth(t)=εΣβpi )' ph+ζΔWrh
(t-]), update amount Δ of weight values between h layer = i layer
W.

Calculate (1). Then, according to the calculated update amount, the weight value W, 1(t) = W for the next update cycle
ji (t-1) ten ΔWJ, (t) Wth (t) = Wl
By repeating the method of determining h(t-1)+ΔW t , D), the output signal y□ from the output layer that is output when the learning presentation signal is presented, and its output signal ypJ The weight values W, , W, and the threshold value θ1.
θ1 will be learned.

In addition, the present applicant previously applied for the “Japanese Patent Application No. 62-3334”.
As disclosed in No. 84 (filed on December 28, 1988, ``Nentwork configuration data processing device''), "1" is always output to the h layer on the input side, and a threshold value θ is weighted on the output. By providing a unit that assigns it as a value, we proposed to incorporate the threshold θ into the weight value W and treat the threshold θ as a weight value.From now on, learning the threshold θ is also similar to learning the weight value W. It will be learned by the learning process.

As shown in FIG. 7, the hierarchical network simulator 30 includes a network structure management section 31 that manages the internal coupling relationship between the input unit 1° and the basic unit 1 of the hierarchical network structure section 20a, and controls the entire arithmetic processing. the calculation control unit 32 and the hierarchical network structure unit 20a
a human data development section 33 that temporarily develops the input values input to the input unit 1' and each unit of the basic unit 1; a weight value management section 34 that manages the weight values assigned to each internal connection; It is configured to include an arithmetic execution section 35 that is provided to simulate the arithmetic function of the unit 1, and an output data management section 36 that manages the arithmetic results of the arithmetic execution section 35. The calculation execution unit 35 includes a multiplication value calculation unit 37 that calculates the multiplication value of the weight value and the input value, a summation value calculation unit 38 that calculates the summation value of the calculation values of the multiplication value calculation unit 37, and It is configured to include a threshold summer peak section 39 that performs dark value processing on the value calculated by the summation value calculation section 38.

With this configuration, the hierarchical network simulator 30 simulates the data processing function executed by the hierarchical network structure section 20a installed in the fuzzy controller. Note that the hierarchical network structure section 2
When configuring 0a by a program means, it can be realized by using such a hierarchical network simulator 30, for example.

When realizing the desired arithmetic function of the antecedent part arithmetic unit 17 according to the hierarchical network structure unit 20a, the user first generates the input/output signal relationship realized by the arithmetic function and then uses the learning signal presentation device 40. A process of registering in the learning signal storage unit 41 is executed. That is, the antecedent part calculation unit 17
If you want to assign an arithmetic function that extracts the minimum truth value (corresponding to the logical product operation) calculated by the antecedent truth value calculation unit 16 as the arithmetic function of the input signal and those input signals, A process is performed in which a human output signal relationship is generated in which the input signal that takes the minimum value among the input signals is used as an output signal, and is registered in the learning signal storage section 41.

FIG. 8 illustrates the input/output signal relationship of this AND operation when the number of input units 1' in the input layer of the hierarchical network structure section 20a is two. This eighth
The example shown in the figure shows an example in which 81 input/output signal relationships are generated.

Next, the user activates the learning signal presentation device 40 by instructing that the input signal in the registered input/output signal relationship be used as a learning presentation signal and the output signal associated with this as a learning teacher signal. At the same time, by activating the learning processing device 50 in response to this activation process, the weight values of the weight value management unit 34 of the hierarchical network simulator 30 enter learning as much as possible to realize this input/output signal relationship. Instruct.

In accordance with this startup instruction, the hierarchical network simulator 30 operates as the hierarchical network structure section 20a installed (or planned to be installed) in the fuzzy controller,
In addition to converting and outputting the presented learning presentation signal according to the data conversion function defined by the weight value, the learning processing device f50 receives the output signal from the hierarchical network simulator 30 and performs the above-mentioned back-up signal. The learning signal presentation device 4 updates the weight value (managed by the weight value management unit 34) of the internal connection according to the propagation method.
0, the learning signal is repeatedly presented until the output signal from the hierarchical network simulator 30 approximately matches the learning teacher signal. Through this process, the weight values of the internal connections of the hierarchical network structure section 20a that realize the registered input/output signal relationship are stored in the weight value management section 34 of the hierarchical network simulator 30.

FIG. 9 shows learning data when learning is performed using the human output signal relationship shown in FIG. 8 as a learning signal. Here, FIG. 9(a) shows the output signal data of the hierarchical network simulator 30 for presentation of each learning presentation signal when the number of learning times is 0 (that is, at the start of learning), and FIG. 9(b) is the output signal data of the hierarchical network simulator 30 for the presentation of each learning presentation signal when the number of learning times is 5000, and FIG. 9(c) is the output signal data for the presentation of each learning presentation signal when the number of learning times is 14673. 3 represents output signal data of the hierarchical network simulator 30. Here, this learning was performed using a hierarchical network structure section 20a whose middle layer consisted of 10 units in one stage.

Furthermore, at the start of learning, a random value generated by a random number generating means was assigned as an initial value of the weight value (including the threshold value) of each internal connection. This figure shows that, for example, when the rAOIJ learning presentation signal of the learning signal in FIG. 8 is given to the hierarchical network simulator 30, the number of learning times is 14673.
This indicates that "0.077" was output from the hierarchical network simulator 30 at the time. Here, "0.100" associated with this "0.077" in parentheses is the learning teacher signal of rAOIJ.

FIG. 10 shows a plot of the number of times of learning and the error of the data processing function of the hierarchical network simulator 30. As shown in this figure, when the number of learning times exceeds 5000, the hierarchical network simulator 30 will output a learning teacher signal roughly corresponding to the presentation of the learning signal shown in FIG. In this state, even if something that is not in the learning signal is presented, it will operate to output an output signal that is close to the content of the AND operation.

FIG. 11 shows learning of weight values and threshold values (managed by the weight value management unit 34 of the hierarchical network simulator 300) assigned to each internal connection of the hierarchical network structure unit 20a when the number of learning times is 14,673. Illustrate the value. Here, "-0,652426" is the weight value of the internal connection between the first unit of the input layer and the first unit of the hidden layer, and "-0,457654" is the weight value of the internal connection between the unit of the output layer and the first unit of the hidden layer. The weight value of the inner connection with the first unit,
"-0,343209" represents the threshold of the first unit of the intermediate layer, and "4.409703" represents the threshold of the unit of the output layer.

As can be seen from Figures 9(c) and 10, this first
The weight value management unit 2 of the hierarchical network structure unit 20a implemented in the fuzzy controller stores the weight values and darkness values shown in FIG.
1a, the hierarchical network structure section 20a
The antecedent part calculation unit 17 that executes the calculation process selects and outputs the minimum truth value calculated by the antecedent part truth value calculation unit 16.

On the other hand, the user performs the calculation function of the antecedent part calculation unit 17 as follows.
The algebraic product of the truth values calculated by the antecedent truth value calculation unit 16 (2 because it is assumed that the number of input units 1 in the input layer of the hierarchical network structure unit 20a is 2)
If you want to allocate an arithmetic function that outputs an algebraic product of two truth values, first, you need to create an input/output signal whose output signal is the input signal and the algebraic product of those input signals, as shown in Figure 12. A relationship is generated and registered in the learning signal storage unit 41, and then the learning signal presentation device 40 and the learning processing device 50
By starting Hierarchical Network Simulator 3
A weight value of 0 in the weight value management unit 34 instructs to start learning as much as possible to realize this input/output signal relationship.

FIG. 13 illustrates learning data when learning is performed using the input/output signal relationship shown in FIG. 12 as a learning signal. Here, this learning is similar to the learning of logical products, where the middle layer of the hierarchical network structure section 20a is composed of
It was made up of units. Further, FIG. 13 shows the output signal data of the hierarchical network simulator 30 when the number of learning times is 9499, at which the learning error has almost converged.

FIG. 14 illustrates the learning values of the weight values and dark values assigned to each internal connection of the hierarchical network structure section 20a at this number of learning times. If the weight values and dark values shown in FIG. 14 are registered in the weight value management section 21a of the hierarchical network structure section 20a installed in the fuzzy controller, the antecedent for executing the arithmetic processing by the hierarchical network structure section 20a is set. The part calculation unit 17 includes the antecedent part truth value calculation unit 1
An antecedent operation is performed in which the algebraic product of the truth values calculated in step 6 is calculated and output.

In addition, the user can perform the calculation function of the antecedent part calculation unit 17 as follows:
The marginal product value Z (hierarchical network structure unit 20a
It is assumed that the number of input units 1 in the input layer of is 2, so if you want to assign an arithmetic function that outputs the marginal product of two truth values, use z-0(x+y- 1) First, as shown in FIG. 15, an input/output signal relationship in which the output signal is the marginal product of two input signals and those input signals is generated and registered in the learning signal storage section 41, and then, By activating the learning signal presentation device 40 and the learning processing device 50, the weight value of the weight value management unit 34 of the hierarchical network simulator 30 instructs to enter learning as much as possible to realize this input/output signal relationship.

FIG. 16 illustrates learning data when learning is performed using the input/output signal relationship shown in FIG. 15 as a learning signal. Here, this learning is similar to the learning of logical products, where the middle layer of the hierarchical network structure section 20a is composed of
It was made up of units. Further, FIG. 16 shows the output signal data of the hierarchical network simulator 30 when the number of learning times is 1628, at which the learning error has almost converged.

FIG. 17 illustrates the learning values of the weight values and dark values assigned to each internal connection of the hierarchical network structure section 20a at this number of learning times. If the weight values and threshold values shown in FIG. 17 are registered in the weight value management unit 21a of the hierarchical network structure unit 20a installed in the fuzzy controller, the antecedent part for which the arithmetic processing is executed by the hierarchical network structure unit 20a is registered. The calculation unit 17 includes an antecedent truth value calculation unit 1
The antecedent part calculation is executed to calculate and output the marginal product of the truth values calculated in step 6.

As described above, according to the present invention, the calculation contents of the antecedent part calculation executed by the antecedent part calculation part 17 can be changed by simply changing the value of the weight value (threshold value) assigned to the internal connection of the hierarchical network structure part 20a. This allows you to freely change and set the settings.

Although the illustrated embodiment has been described, the present invention is not limited thereto. For example, the network structure of the network structure section 20 is not limited to a hierarchical network structure. Further, the calculation processing of the basic unit 1 is not limited to the dark value conversion processing. In addition, the applicant:
In the earlier application, ``Patent Application No. 1983-227825 (filed on September 12, 1988, ``Learning Processing Method for Network Configuration Data Processing Device''), we attempted to improve the back propagation method to shorten the processing time. Although the present invention discloses an invention that enables weight value learning processing to be realized in be able to. Then, the network structure unit 20 is implemented in the fuzzy controller without using a simulator or the like.
It is also possible to use that information to learn the weight values.

(Effects of the Invention) As explained above, according to the present invention, the calculation contents of the antecedent part calculation of the fuzzy controller can be freely controlled without having to prepare a subroutine in advance in the device. Moreover, even with fuzzy controllers that have already been shipped, it is possible to change the calculation contents of this antecedent part calculation without adding or changing hardware or software.From now on, the fuzzy control rules can be changed. Since requests to change the antecedent part operations described in the table can be easily handled, it is possible to improve the performance of the fuzzy controller.

[Brief explanation of drawings]

Figure 1 is a diagram of the principle configuration of the present invention, Figure 2 is an embodiment of the present invention, Figure 3 is a diagram of the basic configuration of the basic unit, Figure 4 is a diagram of the basic configuration of the hierarchical network structure, and Figure 5 is a diagram of the basic configuration of the hierarchical network structure. FIG. 6 is an example of the basic unit; FIG. 6 is an example of the hierarchical network structure; FIG. 7 is an explanatory diagram of the learning system used in the present invention; FIG. 8 is the learning signal generated when assigning logical product operations. 9 and 10 are explanatory diagrams of learning data for logical product operation. Figure 11 is an explanatory diagram of learning data for weight values and dark values for logical product operation. Figure 12 is an explanatory diagram of learning data for logical product operation. Fig. 13 is an explanatory diagram of learning data for algebraic product operations; Fig. 14 is an explanatory diagram of learning data for weight values and threshold values for algebraic product operations; Fig. 15 is a limit diagram. FIG. 16 is an explanatory diagram of a learning signal generated when assigning a product operation, FIG. 16 is an explanatory diagram of learning data for marginal product operation, and FIG. 17 is an explanatory diagram of learning data of weight values and threshold values for marginal product operation. In the figure, 1 is a basic unit, 1' is an input unit, 10 is a fuzzy controller, 16 is an antecedent truth value calculation section, 17 is an antecedent operation section, 1B is a consequent operation section, and 20 is a network structure. 20a is a hierarchical network structure section, 21 is an internal state value management section, 21a is a weight value management section, 30 is a hierarchical network simulator, 40 is a learning signal presentation device, 5
o is a learning processing device.

Claims (3)

[Claims] (1) An antecedent truth value calculation unit (16) that calculates the membership function value for the input control state quantity, and a logical operation on the calculated value of the antecedent truth value calculation unit (16). an antecedent part calculation unit (17) that obtains an output value; a consequent part calculation unit (18) that performs a logical operation on the output value of the antecedent part calculation unit (17) to obtain an output value; A fuzzy controller comprising an inference value calculation unit (19) that calculates and outputs a control operation amount from the output value of the operation unit (18) and a membership function value regarding the control operation amount, the antecedent operation unit (1
7) receives one or more inputs and an internal state value by which the inputs are to be multiplied to obtain a product-sum value, and obtains an output value by performing a function transformation on the product-sum value. A fuzzy controller characterized in that it is constituted by a network structure section (20) constituted by internal connections of basic units (1). (2) In the fuzzy controller according to claim (1), the network structure section (20) is an input unit that distributes the calculated value of the antecedent part truth value calculation section (16) to which the basic unit (1) is input. (1') as a structural unit, a plurality of input units (1'-h) as an input layer, and one or more basic units (1-i) as an intermediate layer. and one basic unit (1-j) as an output layer, between the input layer and the foremost intermediate layer, between the intermediate layers, and between the final intermediate layer and the output layer. A fuzzy controller comprising a hierarchical network structure that constitutes internal connections between which internal state values are set. (3) In the fuzzy controller according to claim (1) or (2), the network structure section (20) has a minimum input signal value as an internal state value assigned to an internal connection of the network structure section (20). A fuzzy controller characterized in that a value is set that causes the controller to operate so as to output.