JPH0668061A - Fuzzy inference device and its operation method - Google Patents

Abstract

PURPOSE:To easily handle information which is difficult to quantize. CONSTITUTION:Input signals expressing information difficult to quantize, for example, R, G, and B signals expressing color information are given to neural networks 3i from an input part 21. Neural networks 3i are preliminarily learnt so that they use teacher samples and teacher data to generate the output signals expressing quantity information. The output signal of the neural network designated by a fuzzy inference rule is given to a fuzzy inference part 12 through a selector 14. Input signals expressing information easy to quantize are given from an input part 11 to the fuzzy inference part 12. The fuzzy inference part 12 uses these input signals to perform the normal fuzzy inference processing in accordance with a preliminarily set rule.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a fuzzy reasoning apparatus and an operating method thereof.

[0002]

2. Description of the Related Art The fuzzy amount that can be handled by the conventional fuzzy inference apparatus has been limited to quantifiable physical amounts such as "slightly large", "reasonably high", and "slightly close". However, in the real world,
A lot of linguistic information that is difficult to quantify is used, such as "yellowish", "pretty", and "interesting".

Therefore, in order to be able to handle linguistic information that is difficult to quantify in the conventional fuzzy inference apparatus, the user must decompose the meaning of the language in advance and define the membership function in detail. I had to leave. For example, it is necessary to consider in advance how much numerical information "yellowish" can be represented by a color signal composed of R (red), G (green), and B (blue) components having a membership function. Need to create.

However, it is very time-consuming and troublesome to define a membership function that represents such ambiguous information that is difficult to quantify.

[0005]

DISCLOSURE OF THE INVENTION It is an object of the present invention to provide a fuzzy inference apparatus that can easily handle ambiguous information that is difficult to quantify and its operating method.

A fuzzy reasoning device according to the present invention is a first fuzzy reasoning device.
A neural network means for receiving an input signal of the above and generating an output signal corresponding to the first input signal in accordance with the neural network structure, and a teacher for giving the neural network means a teacher signal to be the norm of the output signal. A signal generating means, a learning control means for receiving the output signal of the neural network means and the teacher signal of the teacher signal generating means, and modifying the structure of the neural network so that the output signal matches the teacher signal; And a plurality of second input signals including the output signal of the neural network means as at least one input signal, executing fuzzy inference processing according to a plurality of preset fuzzy rules, and at least one Fuzzy reasoning means to generate output signal Eteiru.

It goes without saying that the neural network means may be one or plural.

In a preferred embodiment of the present invention,
A second input signal including an output signal of the neural network means in the inference mode by activating the neural network means, the teacher signal generating means, and the learning controller means in the learning mode to perform learning. Control means for controlling the fuzzy inference means to execute the fuzzy inference processing in response to the above.

Particularly, when a plurality of the neural network means are provided, in order to sequentially perform the operation in the learning mode for each neural network means, the plurality of the neural network means and the neural network to be learned. First selector means for selecting the means are provided.

In order to simplify the configuration of the fuzzy inference means, in the embodiment in which the output signals of the plurality of neural network means are sequentially switched and applied to the fuzzy inference means, the plurality of neural network means,
And fuzzy inference signals to be executed by the fuzzy inference means among the output signals of the plurality of neural network means.
Second selector means is provided for selecting an output signal defined by a rule and giving it to the fuzzy inference means as its second input signal.

The present invention also provides a method of operating the above fuzzy reasoning apparatus. In this operating method, a neural network means for receiving a first input signal and generating an output signal corresponding to the first input signal in accordance with a neural network structure, a reference of the output signal to the neural network means, A teacher signal generating means for giving a teacher signal to be generated, an output signal of the neural network means and a teacher signal of the teacher signal generating means are provided, and the structure of the neural network is such that the output signal matches the teacher signal. A learning controller means for modifying the above, and a fuzzy reasoning means for receiving a plurality of second input signals, executing fuzzy reasoning processing according to a plurality of preset fuzzy rules, and generating at least one output signal In fuzzy reasoning device, in learning mode, The first input signal is applied to the ural network means, a teacher signal corresponding to the first input signal is generated from the teacher signal generating means, and the learning controller means is activated to learn the neural network means. And applying a first input signal to the neural network means in the inference mode and using the output signal of the neural network means as at least one of the second input signals to generate a plurality of second input signals. Is given to the fuzzy inference means to cause the fuzzy inference means to execute fuzzy inference processing according to a plurality of fuzzy rules.

According to the present invention, the learning of the neural network means is performed in the learning mode.
By applying the first input signal to the neural network means and generating a desired teacher signal, the neural network means can quantify the first input signal in a user-friendly or simple form. It is adjusted so as to have a structure for converting into an output signal. In the inference mode, the output signal of the neural network means is input to the fuzzy inference means in the same manner as the general input signal of the fuzzy inference means, and is subjected to the fuzzy inference processing.

As described above, according to the present invention, even an input signal representing information that is difficult to quantify can be given to the neural network means and given a teacher signal in a form desired by the user for learning. Therefore, the learned neural network means has a function of converting input information that is difficult to quantify into a quantified output signal in a form desired by the user. In this way, according to the present invention, information that is difficult to quantify can be easily quantified and handled as an input signal for fuzzy reasoning.
The user will be freed from the troublesome work of setting the numerical value of the membership function as in the past. Moreover, since the form of the teacher signal can be freely selected or decided by the user, the subjectivity of the user can be incorporated into the structure of the neural network. This means that fuzzy reasoning that incorporates the subjectivity of the user is possible.

The present invention further provides a fuzzy inference apparatus equipped with a neural network in which desired functions are set by learning.

The fuzzy inference apparatus receives a first input signal and generates an output signal corresponding to the first input signal according to a neural network structure already set by learning, a neural network means,
And a plurality of second input signals including the output signal of the neural network means as at least one input signal, executing fuzzy inference processing according to a plurality of preset fuzzy rules, and at least one A fuzzy inference means for generating an output signal is provided.

The number of neural network means in this fuzzy inference apparatus may be one or plural.

Particularly when a plurality of the neural network means are provided, the output signal defined by the fuzzy rule to be executed by the fuzzy inference means is selected from the output signals of the plurality of neural network means. When the fuzzy inference means is provided with selector means for supplying the fuzzy inference means as a second input signal thereof, the output signals of a plurality of neural network means are switched to be applied to the fuzzy inference means and the structure of the fuzzy inference means is simplified. To be done.

The present invention further provides a method of operating the above fuzzy reasoning apparatus. The method of operation comprises a neural network means for receiving a first input signal and generating an output signal corresponding to the first input signal according to a neural network structure already set by learning, and a plurality of second network signals. In a fuzzy inference apparatus having fuzzy inference means for receiving an input signal, executing a fuzzy inference process according to a plurality of preset fuzzy rules, and generating at least one output signal, the neural network means is provided with A plurality of second input signals to the fuzzy inference means by using the output signal of the neural network means as at least one of the second input signals and the plurality of second input signals to the fuzzy inference means. Fuzzy
It executes fuzzy inference processing according to rules.

Also in the fuzzy inference apparatus of the present invention,
By making the neural network means perform appropriate learning, it becomes possible to handle linguistic information that is difficult to quantify in fuzzy reasoning processing, and is freed from the troublesome work of setting the numerical value of the membership function.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram showing an electrical configuration of a fuzzy reasoning apparatus.

This fuzzy inference apparatus has a learning mode and an inference mode. The operation in the learning mode is performed before the operation in the inference mode. In the learning mode, the input data (input signal) is separately supplied from the input unit 21 to each of the plurality of neural networks 31 to 3n included in the neural network group 30, and the teacher data (teacher signal) is supplied from the teacher data input unit 25. Is given and learning is performed so that the outputs of the neural networks 31 to 3n match the corresponding teacher data. In the inference mode, according to the fuzzy inference rules and the inference rules stored in the membership function storage unit 13, the fuzzy inference unit 12 uses the input signals provided from the input unit 11 and the neural network group 30 to perform fuzzy inference. Run.

The fuzzy inference unit 12 can be realized by a hardware architecture, or by a computer programmed to execute processing according to fuzzy rules. It is also possible to realize a part by hardware and another part by software. However, one computer system in which all of the fuzzy inference unit 12, selector 14, overall control unit 17, neural network control unit 22, selector 23, learning controller 24 and neural network group 30 shown in FIG. 1 are programmed. It is also possible to build by.

The inference rule and membership function storage unit 13 is composed of a writable memory, and is used by a plurality of fuzzy inference rules by which the fuzzy inference unit 12 should execute fuzzy inference processing according to it and these fuzzy inference rules. The data representing the membership function is stored in advance. Depending on the configuration of the fuzzy inference unit 12, for example, when the fuzzy inference unit 12 has an architecture dedicated to fuzzy inference processing, at least one or all of the fuzzy inference rules and / or membership functions may be partially or wholly. It may be incorporated as the architecture of.

The input section 11 provides an input signal (input data) to the fuzzy inference section 12, and is generally composed of various sensors and a circuit for processing signals from the sensors. A part of the input unit 11 may be composed of a keyboard or various data processing circuits. Input section 11 in general
From, a plurality of types of input signals are given to the fuzzy inference unit 12.

The output unit 15 converts a signal representing the inference result obtained from the fuzzy inference unit 12 into a signal suitable for an external device (control target, actuator, etc.). The output unit 15 may include a display device or a printer. Even if the defuzzifier is included in the fuzzy inference unit 12, the output unit
Either may be included in 15.

The neural network group 30 includes a plurality of neural networks 31-3n. These neural networks 31-3n will typically be implemented by a computer system that includes memory.

The input unit 21 is a neural network group 30.
The input signal (input data) is given to. Whereas the input unit 11 described above provides a kind of input that is easy to quantify and relatively easy to create a membership function, the input unit 21 is difficult to quantify or creates a membership function. This is a complicated type of input. An example of the input unit 21 is a television / camera that outputs a video signal as described later. The same kind of input signal may be given to each of the plurality of neural networks 31 to 3n, or an entirely different kind of input signal may be given.

The teacher data input unit 25 provides teacher data (teacher signal) to each of the plurality of neural networks 31 to 3n in the learning mode, and may be composed of an input device such as a keyboard and a mouse.

In the learning mode, the learning controller 24 inputs the output signal of the neural network and the corresponding teacher data given from the teacher data input unit 25 for each of the neural networks 31 to 3n, The structure of the neural network is adjusted so that the output signal matches the teacher data. This learning controller 24 is also realized by a computer system.

The neural network control unit 22 uses the selector 23 and the learning controller in the learning mode.
24 is controlled to perform the learning operation. The selector 23 selects one neural network as a learning target. Neural network 31
Neurals selected by selector 23 out of ~ 3n
The network is enabled. However, the selector
23 may be provided between the output side of the neural networks 31 to 3n and the learning controller 24.

The selector 14 selects one of the output signals of the plurality of neural networks 31 to 3n in the inference mode and supplies it to the fuzzy inference unit 12. The output signal of the neural network selected by the selector 14 is determined by the fuzzy inference rule to be processed by the fuzzy inference unit 12 accordingly. However, if the antecedent processing unit corresponding to each of the neural networks 31 to 3n is provided in the fuzzy inference unit 12, it is not necessary to provide the selector 14. selector
Providing 14 simplifies the configuration of the antecedent processing unit in the fuzzy inference unit 12.

The control input unit 16 sets a learning mode or an inference mode, and is realized by a keyboard or a switch, for example. The mode specifying signal input from the control input unit 16 is given to the overall control unit 17. In response to the mode designating signal, the general control unit 17 activates the neural network control unit 22 in the learning mode to perform the operation in the learning mode, and in the inference mode, the fuzzy inference unit 12 is activated. Inference processing will be executed.

FIG. 2 shows the structure of a general neural network. This neural network is composed of an input layer 40, an intermediate layer 50 and an output layer 60, and each layer is composed of a plurality of neurons 41-4k, 51-5l, 61-6m. The neurons of the input layer 40, the intermediate layer 50, and the output layer 60 are connected to each other by a synaptic weight. The output of each neuron can be determined to take a binary value of 0 or 1, a multivalued value, a continuous value, or the like.

The input data given to the neural network is inputted to the corresponding neurons 41 to 4k of the input layer 40. The neurons 41 to 4k of the input layer 40 output the input data as they are to all the neurons 51 to 5l of the intermediate layer 50. The output data of the input layer 40 is weighted and input to each of the neurons 51 to 5l of the intermediate layer 50. For example, when the neuron output in the middle layer is binary,
The output is determined by performing threshold processing on the sum of weighted data of the input data given from all the neurons 41 to 4k of the input layer 50. Furthermore, the output data from the neurons 51 to 5l of the intermediate layer 50 are also weighted and input to all the neurons 61 to 6m of the output layer 60. The same processing as in the neurons of the intermediate layer 50 is performed in each neuron of the output layer 60. Eventually, the input data propagates through each layer of the neural network and is output as output data.

Learning in such a neural network is performed as follows. A teacher sample is input as input data from the input unit 21 to the neural network. Teacher data corresponding to this teacher sample is given from the teacher data input unit 25 to the learning controller 24. When a teacher sample is given as input data, it propagates through each layer 40-60 of the neural network and outputs output data corresponding to the input teacher sample. The learning controller 24 compares this output data with the teacher data. When the output data of the neural network and the teacher data are different, the learning controller 24 adjusts the above-mentioned weight in each neuron of the neural network so as to reduce the error. The teacher data corresponding to the teacher sample as the input data is repeatedly given until the error becomes equal to or less than the predetermined value, and the learning is continued.

When, for example, an error back-propagation learning algorithm (generalized delta rule) is used as a learning algorithm, a teacher sample is first given to the network, which propagates forward through the network to obtain output data. The difference between the output data and the teacher data is multiplied by the differential coefficient to calculate the error for the neuron in the output layer of the network. Next, the error signal propagates backward through the network, and the error signal for each neuron is calculated, and the weights are corrected based on this. Other algorithms may be used as the learning algorithm.

The one selected by the selector 23 for each of the plurality of neural networks 31 to 3n shown in FIG. 1 is sequentially subject to learning, and learning is performed for all neural networks.

Even if an input signal representing information that is difficult to quantify is given to the neural network, the neural network trained by giving appropriate teaching data
If it is a network, corresponding information that can be easily quantified can be obtained from the neural network.

Consider, for example, the following fuzzy knowledge (inference rule).

(Rule 1) IF curvature is "large" & area is "small" & color is "yellowish", and it is THEN product A

Regarding the curvature and area of the antecedent part in the above fuzzy inference rule, "large" and "small"
It is easy to express the information as a numerical value. Therefore, the input about the curvature and the area is given from the input unit 11.

However, the "yellowish" information about color is sensuous and it is difficult to express this numerically.

Sensitive information such as color can be easily converted into a quantity including the subjectivity of the user by giving an appropriate teacher sample and teacher data to the neural network.

FIG. 3 shows an example of a structure for handling information about colors. The R, G, B color signal components input from the television camera 21A are applied to the neural network. The output layer has 2 bits. The teacher data is shown in FIG. The degree of yellowness is "true yellow", "yellowish", "yellowish" and "not yellow"
There are four stages, teacher data "11", "1"
"0", "01" and "00" are associated with each other.
The color samples R, G, and G obtained by making the camera 21A pick up the teacher sample corresponding to the above-described four levels of yellow degree are obtained.
The neural network 31 is trained by inputting B into the neural network and giving corresponding teacher data.

By such learning, it is difficult to quantify,
The color information, that is, the membership functions for R, G, and B that had to be created in the conventional case, is quantified and output as the output signal of the 2-bit neural network. . Moreover, since the teacher sample and the teacher data in the learning of the neural network can be given by reflecting the user's subjectivity, it is possible to obtain the output data reflecting the user's subjectivity.

FIG. 5 shows a part of the configuration of the fuzzy inference unit 12 (particularly the antecedent processing unit). Goodness of fit calculator 7
For example, the above-mentioned rule 1 and the membership function corresponding thereto are given to the 1, 72, and 74 from the storage unit 13 and set. Input data on curvature and area are given from the input unit 11 to the fitness calculation units 71 and 72.

The code representing the color proposition set in the above-mentioned rule 1 is determined by the selector control unit 79. Assuming that the neural network learned about the above-mentioned degree of yellow is the neural network 31, the control unit 79 controls the selector 14 so as to select the output signal of the neural network 31. on the other hand,
The television camera 21A photographs an object, outputs a color component signal representing the object, and this color component signal is given to the neural network 31, so that the neural network 31 indicates the yellow degree of the object. Generates 2-bit output data. This output data is given to the fitness calculating section 74 via the selector 14.

FIGS. 6A, 6B and 6C show membership functions of curvature, area and yellowness, respectively. Three types of membership functions are set for the curvature and area: "small", "medium", and "large". As for the degree of yellowness, as shown in FIG. 4, four kinds of membership functions of "true yellow", "yellowish", "yellowish" and "not yellow" are set.

Given an input signal for curvature,
The membership function "large" in the fitness calculation unit 71
The fitness of the input signal to Similarly, for the area, the fitness calculating unit 72 obtains the fitness of the input signal with respect to the membership function “small”.

Regarding the color, the output data of the neural network 31 indicating the degree of yellowness is given to the fitness calculating section 74 via the selector 14 and the fitness of this output data to the membership function "yellowish" is obtained.

The MIN calculation unit 78 performs MIN calculation on the data indicating the conformity of the curvature, area and yellowness output from the conformity calculation units 71, 72 and 74, and outputs the antecedent part calculation output indicating the peculiarity of the product A. Will be obtained.

Although not shown, the fuzzy inference unit
Reference numeral 12 denotes an arithmetic circuit (for example, a MAX arithmetic unit) between the antecedent MIN arithmetic result and the consequent membership function (or singleton) of each fuzzy inference rule, and an arithmetic circuit for integrating inference outputs of a plurality of rules. Needless to say, a defuzzifier and the like are provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing an electrical configuration of a fuzzy reasoning apparatus.

FIG. 2 is a block diagram showing a configuration of a neural network.

FIG. 3 is a block diagram showing an input unit and a neural network.

FIG. 4 shows an example of teacher data.

FIG. 5 is a block diagram showing a part of a configuration of a fuzzy inference unit.

FIG. 6 (A) is a membership function for curvature,
(B) is a graph showing the membership function for area, and (C) is a graph showing the membership function for yellowness.

[Explanation of symbols]

 11, 21 Input part 12 Fuzzy inference part 13 Inference rule and membership function storage part 14, 23 Selector 15 Output part 16 Control input part 17 Overall control part 21A TV camera 22 Neural network control part 24 Learning controller 25 Teacher data input Part 30 Neural network group 31-3n Neural network

Claims (8)

[Claims] 1. A neural network for receiving a first input signal,
Neural network means for generating an output signal corresponding to the first input signal according to a network structure, teacher signal generating means for giving the neural network means a teacher signal to be a norm of the output signal,
A learning control means for modifying the structure of the neural network so that the output signal of the neural network means and the teacher signal of the teacher signal generating means are given and the output signal matches the teacher signal, and the neural network means. Accept a plurality of second input signals including the output signal of the network means as at least one input signal, execute fuzzy inference processing according to a plurality of preset fuzzy rules, and generate at least one output signal A fuzzy inference device equipped with a fuzzy inference means. 2. In the learning mode, the neural network means, the teacher signal generating means and the learning controller means are activated to cause the neural network means to perform learning, and the output signal of the neural network means is included in the inference mode. The control means for controlling to cause the fuzzy inference means to execute fuzzy inference processing in response to the second input signal.
The fuzzy inference device described in. 3. The fuzzy inference apparatus according to claim 1, further comprising a plurality of the neural network means and a first selector means for selecting a neural network means to be learned. 4. The fuzzy inference by selecting an output signal defined by a fuzzy rule to be executed by the fuzzy inference means from among output signals of the plurality of neural network means and the plurality of neural network means. The fuzzy inference apparatus according to claim 1, further comprising second selector means for giving the means as its second input signal. 5. Neural network means for receiving a first input signal and generating an output signal corresponding to said first input signal according to a neural network structure already set by learning, and said neural network.
Accept a plurality of second input signals including the output signal of the network means as at least one input signal, execute fuzzy inference processing according to a plurality of preset fuzzy rules, and generate at least one output signal A fuzzy inference device equipped with a fuzzy inference means. 6. The fuzzy inference by selecting an output signal defined by a fuzzy rule to be executed by the fuzzy inference means from among output signals of the plurality of neural network means and the plurality of neural network means. 6. A fuzzy inference apparatus according to claim 5, further comprising selector means for giving the means as its second input signal. 7. A neural input for receiving a first input signal
Neural network means for generating an output signal corresponding to the first input signal according to a network structure, teacher signal generating means for giving the neural network means a teacher signal to be a norm of the output signal,
Learning controller means for applying an output signal of the neural network means and a teacher signal of the teacher signal generating means, and modifying the structure of the neural network so that the output signal matches the teacher signal, and a plurality of learning controller means. In a learning mode, in a fuzzy inference device having fuzzy inference means for accepting two input signals, executing fuzzy inference processing according to a plurality of preset fuzzy rules, and generating at least one output signal, in the learning mode, Applying the first input signal to the neural network means, and generating a teacher signal corresponding to the first input signal from the teacher signal generating means,
The learning controller means is activated to activate the neural network.
In the inference mode, the network means is trained and a first input signal is provided to the neural network means, and the output signal of the neural network means is used as at least one of the second input signals to generate a plurality of first input signals. A method of operating a fuzzy inference apparatus, wherein the fuzzy inference means is caused to execute fuzzy inference processing according to a plurality of fuzzy rules by giving the input signal 2 to the fuzzy inference means. 8. A neural network means for receiving a first input signal and generating an output signal corresponding to the first input signal according to a neural network structure already set by learning, and a plurality of second network signals. Accepts an input signal and presets multiple fuzzy
In a fuzzy inference apparatus including a fuzzy inference means for executing a fuzzy inference process according to a rule and generating at least one output signal, the neural network means is provided with a first input signal and the neural network means is provided. The second input signal as at least one of the second input signals to the fuzzy inference means to cause the fuzzy inference means to execute fuzzy inference processing according to a plurality of fuzzy rules. Operation method of fuzzy reasoning device.