JPH03225427A - Fuzzy reasoning system - Google Patents

Abstract

PURPOSE:To evade an increase in system scale by providing a rule fitness neural network which outputs a fitness vector and an inference output neural network which receives an input vector and a rule fitness vector and outputs an inference result. CONSTITUTION:This system is equipped with the rule fidelity neural network 4 which inputs input data and outputs fitness for each rule of the input data and the inference output neural network 6 which inputs the input data and rule fitness and outputs a inferred value. A value indicating how much the input data matches each rule is used as data for learning of this neural network 4 by using a fuzzy clustering method. Further, an evaluation norm used to determine the number of rules is predetermined by an evaluation norm setting neural network 15 which inputs the input data and the inferred value for the input data and outputs the name of the label of the evaluation norm suitable to their combination. Consequently, the fuzzy inference system using the small- norm neural networks is obtained.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to an ambiguous inference system, and in particular to an ambiguous inference system.
Concerning fuzzy reasoning systems that draw conclusions from uncertain information and knowledge.

[Conventional technology]

Fuzzy inference systems using fuzzy set theory exist as conventional fuzzy inference systems that draw conclusions from ambiguous and uncertain information and knowledge. In that reasoning, M
In the inference rule called the production rule "F (antecedent part) THEN (consequent part)", "large"
, "small" and other vague values that cannot be clearly defined can be described using fuzzy sets.

FIG. 4 is a diagram showing the grades of feature amounts for membership functions.

The description ``large'' can be written, for example, as shown in FIG. 4 using the membership function of fuzzy set theory. However, the horizontal axis is a value obtained by normalizing the physical quantity handled by the maximum value, and the vertical axis is a grade, which is a quantity representing the degree of "largeness". In conventional fuzzy inference systems, the shape of membership and the division of the variable space of rules are determined using heuristics based on the experience of experts.

As a method to solve this problem, a fuzzy modeling method that applies nonlinear programming (Proceedings of the Society of Instrument and Control Engineers, published by the Society of Instrument and Control Engineers in June 1986, Vol.
l, 23. No, 6°pp, 106-108. Sugano's description (hereinafter referred to as method 1) and the method of formulating fuzzy models by explaining nonlinear problems using neural networks (published in May 1988 by the International Fuzzy Systems Society 9 Japan Skin Division Secretariat) Proceedings of the 4th Fuzzy Systems Symposium, pp. 55-
60. Hayashi and Takagi (hereinafter referred to as Method 2) have been published.

[Problem to be solved by the invention]

However, conventional fuzzy inference systems that incorporate such techniques have the following problems.

That is, method 1 has problems such as the amount of processing is large due to the use of nonlinear programming, the evaluation criterion suitable for the model must be selected empirically, and the algorithm is complicated.

Furthermore, since method 2 requires as many neural networks as the number of fuzzy rules, it has the problem that the scale of the system increases if a large number of rules are required.

[Means to solve the problem]

A fuzzy inference system of a first invention is a fuzzy inference system that obtains one or more inference results from one or more input data using a neural network, and includes an input vector input unit that inputs an input vector as the input data; , a rule suitability neural network that outputs a rule suitability vector indicating how well the input vector matches each rule; and one or more sets of inferences in response to the input vector and the rule suitability vector. The inference output neural network outputs the results.

Further, in the fuzzy inference system of the first invention, the fuzzy inference system of the second invention causes the rule suitability neural network to find an appropriate rule suitability vector by applying a fuzzy clustering method to the input vector. It is configured with a fuzzy clustering section.

On the other hand, the third invention's fuzzy inference system is configured by determining the number of rules for the rule suitability of the rule suitability neural network in the first fuzzy inference system using a preset evaluation criterion.

On the other hand, in the fuzzy inference system of the first invention, the fuzzy inference system of the fourth invention inputs an input vector and the inference result to be obtained by the input vector, and outputs a label name of a suitable evaluation criterion. It is configured with an evaluation norm setting neural network.

Furthermore, in the fuzzy inference system of the first invention, the degree of the rule suitability vector output from the rule suitability neural network is set as the number of rules, and the value to which the input vector belongs to each rule is set as the number of rules. It is composed of degrees.

Note that the ambiguous inference system of the sixth invention is configured by realizing one rule suitability neural network and one inference output neural network in the ambiguous inference system of the first invention.

[Effect]

The structure of the fuzzy inference system of the present invention includes a neural network A that inputs input data and outputs the degree of conformance of the input data to each rule, and a neural network A that inputs the input data and the degree of conformity of each rule and outputs an inference value. It consists of two networks: network B. Furthermore, the learning data for the neural network A is assumed to be a value indicating how well the input data conforms to each rule by using a fuzzy clustering method. Furthermore, the evaluation norms used to determine the number of rules are determined in advance by a neural network C that inputs input data and inferred values for the input data and outputs the label name of the evaluation norm suitable for the combination. .

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIGS. 1, 2, and 3 are block diagrams showing an embodiment of the fuzzy inference system of the present invention.

However, FIG. 1 is a block diagram of the learning mode, FIG. 2 is a block diagram of the inference mode, and FIG. 3 is a block diagram of the evaluation norm setting mode.

First, the learning mode will be explained using FIG. 1.

Input vectors, which are one or more input data obtained from the learning input vector input section 1, are fuzzy clustered in the fuzzy clustering section 2 using the number of rules stored in the rule number storage section 3.

Next, the input vector and the rule suitability vector [order is the number of rules] which is the result of fuzzy clustering of the vector are sequentially sent to the rule suitability neural network 4, and when each input vector is input, Learning is performed so that a rule suitability vector for an input vector is output. In parallel, the input vector, rule suitability vector, and learning inference value vector The inference value vector that is the inference result for each input vector obtained from the input unit 5 is sequentially sent to the inference output neural network 6, and each input vector is When the vector and the rule suitability vector are input, learning is performed so that an inference value vector is output.

When the above learning is completed, each output of the inference output neural network 6 and each inference value vector obtained from the learning inference value vector input unit 5 are sent to the evaluation standard value meter nose unit 7, and the evaluation standard value is Calculated. Further, the new standard value, which is the obtained evaluation standard value, is sent to the rule number updating section 9 and compared with the old standard value, which is the value stored in the evaluation standard value storage section 8.

If the new norm value is smaller, the new norm value is stored in the evaluation norm value storage unit 8, the number of rules in the rule number storage unit 3 is increased by 1, and the same learning as described above is performed in the rule suitability neural network 4. and is performed on the inference output neural network 6.

On the other hand, if the new reference value is larger, the rule suitability neural network 4 and the inference output neural network 6 determined in the previous learning are used as the learning results.
Finish learning. However, the initial values of the rule number storage section 3 and the standard value storage section 8 are respectively 2 and infinity.

Next, the inference mode will be explained using FIG. 2.

The input vector obtained from the inference input vector input unit 10 is sent to the rule suitability neural network 11, and a suitability vector is obtained as its output.

Furthermore, the input vector and the fitness vector are sent to the inference output neural network 12, and the inference value vector obtained as its output becomes the inference result. however,
It is assumed that the rule suitability neural network 11 and the inference output neural network 12 are the same as the rule suitability neural network 4 and the inference output neural network 6 obtained in the learning mode.

Next, the evaluation standard setting mode will be explained using FIG. 3. Several input-output relationship models and evaluation norms suitable for the models are determined in advance, and the evaluation norm setting neural network 15 receives the input values and output values of the model as input, and receives the evaluation norms as output. Learning is performed in such a way that when the input and output values of a certain model are input with labels corresponding to each other, and when an evaluation norm suitable for that model is input, the confidence level for the label of the evaluation norm suitable for that model increases.

Then, the input vectors obtained respectively by the norm setting vector input unit 13 and the inference value vector input unit 14 and the preferable inference value vectors for the input vectors are sequentially input to the evaluation norm setting neural network 15.
sent to.

As a result, the reliability of the label of each evaluation standard obtained is sequentially obtained, and the evaluation standard for the label with the largest average value is set as the evaluation standard used in the learning mode, that is, the evaluation standard value meter nose part 8.

Note that the fuzzy clustering method in the fuzzy clustering unit 2 and the neural networks 4, 6, 1
1.12 structure and evaluation reference value Several evaluation reference values for the meter nose section 8 have been proposed in the past, and the present invention is not limited to these regulations.

For example, regarding fuzzy clustering,
Regarding the FCM method and neural network described on pages 171 to 197 of "Practical Application and Application Examples of Fuzzy Systems" published by Triceps in 2010,
September 1989: A method for performing pack propagation learning on a multilayer network, and evaluation reference values, described on pages 28 to 84 of "Introduction and Practical Neurocomputer" published by Gijutsu Hyoronsha. For details, please refer to ``Mathematical Science N'' published by Science Company in March 1976.
o, 153 Special Feature Information Amount Standards'', pages 5 to 11 include A and C, etc.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to realize a fuzzy inference system using a neural network that does not depend on the number of production rules required for inference, and even when the number of rules increases, it is possible to provide a fuzzy inference system using a small-scale network. There is an effect.

[Brief explanation of drawings]

FIG. 1, FIG. 2, and FIG. 3 are block diagrams showing an embodiment of the fuzzy inference system of the present invention, and FIG. 4 is a diagram showing grades of feature amounts for membership functions. DESCRIPTION OF SYMBOLS 1... Learning input vector input unit, 2... Fuzzy clustering unit, 3... Rule number storage unit, 4...
Rule suitability neural network, 5... Learning inference value vector input unit, 6... Inference output neural network, 7... Evaluation norm calculation unit, 8... Evaluation norm storage unit, 9... Rules number updating unit, 10... human vector input unit for inference, 11... rule suitability neural network, 12... inference output neural network, 13... vector input unit for norm setting, 14.
...Inference value vector input section, 15...Evaluation norm setting neural network.

Claims (1)

[Scope of Claims] 1. In a fuzzy inference system that obtains one or more inference results from one or more input data using a neural network, an input vector input unit that inputs an input vector as the input data; a rule suitability neural network that outputs a suitability vector indicating how well the input vector matches each rule;
An ambiguous inference system comprising: an inference output neural network that receives the input vector and the rule suitability vector and outputs one or more sets of inference results. 2. The fuzzy inference system according to claim 1, further comprising a fuzzy clustering unit that causes the rule suitability neural network to find an appropriate rule suitability vector by applying a fuzzy clustering method to the input vector. 3. The ambiguous inference system according to claim 1, wherein the number of rules for the rule suitability of the rule suitability neural network is determined using a preset evaluation criterion. 4. The ambiguous inference system according to claim 1, further comprising an evaluation norm setting neural network that inputs an input vector and an inference result to be obtained from the input vector and outputs a label name of an appropriate evaluation norm. . 5. The ambiguous inference system according to claim 1, wherein the degree of the rule suitability vector outputted from the rule suitability neural network is set as the number of rules, and the value is set as the degree to which the input vector belongs to each rule. 6. The ambiguous inference system according to claim 1, wherein the rule suitability neural network and the inference output neural network are each realized by one.