JPH05100859A - Fuzzy inference device provided with inference control mechanism, and its leaning method - Google Patents

Abstract

PURPOSE:To execute minute inference even if an inference object has complicated structure by providing an inference control mechanism for fuzzy inference, judging the state of the inference object and executing inference through the use of the form and the rule of an optimum membership function. CONSTITUTION:Mechanisms from input 101 to output 109 execute regular fuzzy inference. The inference control mechanism 110 controls a membership function storage device 103 and a rule weight storage device 107 by input from an antecedent part proposition fitting evaluation device 102 and a rule antecedent part fitting evaluation device 104. Namely, the inference control mechanism 110 receives an antecedent part proposition fitting and a rule antecedent part fitting as input. An inference control part calculates the form of the optimum membership function and rule weight by the fuzzy inference system and updates the form parameter of the membership function stored in the membership storage device and rule weight stored in a rule weight reference device. Thus, minute inference becomes possible even if there is only fragmentary knowledge with respect to the inference object.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fuzzy inference method, in which the state of an inference target is determined from the internal state of an inference device in the inference process, and membership functions and rules defined in more detail according to the determined state. Is used to make precise inferences. Therefore, it is particularly effective when the inference target has a plurality of different states or when a complicated state change occurs.

The present invention also provides a reasoning target when the reasoning target has a plurality of different states, when a complicated state change occurs, and when such a state cannot be explicitly expressed in advance in the form of knowledge. Is a fuzzy inference learning device that automatically learns the state of and enables the above fuzzy inference method.
This is effective when the state of the inference target cannot be easily identified.

Further, the present invention relates to automatic knowledge acquisition, and has an automatic knowledge acquisition function for generating a new rule by analyzing a learning result in an inference device which automatically learns a state of an inference target and performs fuzzy inference. This is effective when there is only partial knowledge of the inference target or when there is uncertain knowledge.

[0004]

2. Description of the Related Art When an inference target has several states, a membership function and rule are created for each state in the prior art. At the time of fuzzy inference, the state of the inference target is determined, and the membership function and rule are switched according to the determined state. A production system or a fuzzy inference system is used as a mechanism for judging the state of the inference target and switching membership functions and rules. In this production system and fuzzy inference system, membership functions and rules (metal rules) for switching rules are described. For example, Japanese Patent Laid-Open No. 1-113574 proposes a method of correcting the shape of the membership function according to an input using a fuzzy inference system.

Regarding learning, some methods of preparing input / output data and adjusting fuzzy inference parameters have been proposed. A method for automatically adjusting the rule weights is as follows: 1990 International Joint Neural Network Society Proceedings Vol. 2 55
Pages 58 to 58 (Proceedings of International Joint Con
ference on Fuzzy Logic & Neural Networks).
Also, regarding the method of automatically adjusting the shape of the membership function, the Society of Instrument and Control Engineers, Vol. 24, No. 2 (1984
Year) Discussed on pages 191 to 197. As an automatic knowledge acquisition method of fuzzy reasoning, there is a method of using a statistical method such as pp. 645 to 648, etc.

[0006]

In the conventional membership function and rule switching methods, the inference state is determined directly from the inference input. The state of the inference target is determined only when the inference system receives an input, and the fitness of the proposition, the fitness of the rule, and the inference output obtained in the inference process are not reflected. As a result, if the inference proceeds under the determined state, even if there is a contradiction in the rules, the inference must be continued with the contradiction left. The present invention determines the state of the inference target including not only the input of the inference but also the goodness of fit obtained in the process of inference, the inference output, etc., and switches the membership function and the rule in accordance with the state to make a consistent inference. It is an object of the present invention to provide a fuzzy reasoning device for performing.

The learning by the prior art is limited to the adjustment of the rule weight and the shape of the membership function.
The state classification itself of the inference target is not learned. In the prior art, the state classification of the inference target was empirically determined by a human, and then a membership function and a meta-rule for switching the rules were created. Therefore, when the inference target has many states, it is difficult to determine each state, which makes it more difficult to determine the metarule. This is a serious problem especially when only fragmentary knowledge can be obtained for the inference target. Another object of the present invention is to provide a learning method in which each state of an inference target can be automatically recognized by learning and a metarule is automatically created.

Further, in the prior art, when the inference target is composed of various states, it is necessary to increase the number of divisions of the states to make the inference rules precise in order to perform accurate inference. However, in this case, there is a problem that expansion of rules by acquisition of new knowledge and management of consistency of each rule become complicated. By learning, the shape of the membership function and the rules are paired with each state of the inference target to facilitate the expansion and management of the rules, and to provide a learning method that combines hidden learning with learning. It is still another object of the present invention.

[0009]

In order to achieve the above object, according to the present invention, the input of a reasoning control mechanism provided in a fuzzy reasoning apparatus, the input of fuzzy reasoning, the degree of conformity obtained in the process of fuzzy reasoning, the reasoning output, etc. To use. The inference control target determines the state of the inference target, and sets the shape of the membership function, the weight of the rule, etc. that are optimal for that state.

In order to achieve the above-mentioned other object, in the present invention, the error of each fitness in the inference process is calculated from the error of the inference output, and the inference control mechanism is learned by using this as a teacher signal.

In order to achieve the above-mentioned other object, in the present invention, the internal state of the inference control mechanism is made to correspond to each state of the inference target, a set is made up of each state, the membership function and the rule, and this is used as the inference rule. And Inference rules can be transformed into membership functions and rule expressions as needed.

[0012]

With respect to the above object, the fuzzy inference control mechanism is provided to determine the state of the inference target and control the fuzzy inference according to the determined state of the inference target.

When there is an input to the fuzzy inference apparatus, first, normal fuzzy inference is performed. Next, the inference control mechanism receives as inputs the fuzzy inference input, the goodness of fit of the fuzzy inference process, and the inference output. The target state is the input pattern,
Since it appears characteristically in the antecedent part proposition conformity, the pattern of rule conformity, etc., the state of the inference target can be judged from these. The inference control mechanism is a production system, neural network, fuzzy inference system, etc. for recognizing patterns and determining the state of the inference target, and the shape of the membership function that is optimal according to the state of the inference target (recognized pattern). , Fuzzy reasoning parameters such as rule weights are reset in the fuzzy reasoning device. The fuzzy inference apparatus in which the inference parameters have been reset performs inference again to obtain an inference output. This enables precise inference according to the state of the inference target.

For other purposes, the internal structure of the inference control mechanism is automatically determined by learning, the state of the inference target is determined, and the shape, rule, etc. of the membership function optimal for that state can be output. .. When fuzzy inference is performed, the antecedent part proposition conformity, the rule antecedent part conformity, and the rule conformance are calculated from the input, and the inference output is finally obtained. At the time of learning, the error between the inference output and the desired output is obtained, and this error is calculated in the opposite direction to that at the time of inference to obtain the error of each rule conformity, the error of the rule antecedent part conformity, the error of the proposition, and the like. Since the error of the output of the inference control mechanism is determined by this, the inference control mechanism is learned by using this error as a teacher signal. By learning, the inference control mechanism outputs the shape of the membership function, the rule weight, etc. for which the error in the inference output is the smallest for the input having a certain characteristic.

With respect to the above still other object, it is investigated what kind of input the output of the learned inference control mechanism outputs, and what kind of membership function is applied to what state of the inference target. Make sure that the shapes and rules of are compatible. After learning, the inference control mechanism outputs an optimum output for an input pattern having a certain characteristic. This input pattern corresponds to the state to be inferred, and the output is the shape of the optimum membership function for that state, the weight of the rule, etc. This correspondence is a new rule.

[0016]

EXAMPLES Examples of the present invention will be described below.

A first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of a fuzzy inference system which is the object of the present invention. The dotted line shows the normal fuzzy reasoning procedure. Although there are several different definitions of the fuzzy inference method, the most general method here is, for example, Information Processing Vol. 30, No. 8, pages 942 to 947 (198).
The method described in (9 years) will be used as an example. Of course, this does not limit the scope of the present invention.

A membership function storage device 103 stores shape parameters specifying the shape of the membership function. 102 is an antecedent part proposition conformity calculation device, which determines a membership function from the shape parameters stored in the membership function storage device 103 by referring to the rules, and together with the inference input 101, Calculate the goodness of fit. Reference numeral 104 is a rule antecedent part conformance calculating device, which calculates each rule antecedent part conformance by min operation of the antecedent part proposition conformance. A consequent part proposition conformity calculation device 105 reads the shape parameter of the membership function from the membership function storage device and determines the shape of the membership function. 106
Is a rule conformance calculation device, which calculates the conformance of each rule by multiplying the antecedent conformity of the rule by the consequent proposition conformity and the rule weight stored in the rule weight storage device 107. Reference numeral 108 is an output arithmetic unit, and the degree of conformance to each rule is max
The output of 109 is determined by taking the center of gravity of the combined membership function.

An inference control mechanism 110 receives the antecedent part proposition conformity and the rule antecedent part conformance as inputs. The inference control unit calculates the optimum shape and rule weight of the membership function by the fuzzy inference system, and calculates the shape parameter of the membership function stored in the membership storage device and the rule weight stored in the rule weight storage device. change.

Fuzzy reasoning in a system including a reasoning control mechanism is shown in FIG. First, the proposition conformity of each antecedent part is calculated from the input. Next, in the inference control mechanism, the fuzzy inference system is used to obtain the optimum shape of the membership function from the values of the antecedent part proposition fitness, and the shape parameters of the membership function stored in the membership function storage device are changed. To do. The membership function after the shape change is used to recalculate the antecedent part proposition conformity, and this is used to calculate the rule antecedent part conformance. Again, in the inference control mechanism, the rule weight is calculated from the rule antecedence conformance, and the rule conformance is obtained from the rule weight, the rule antecedence conformance, and the consequent proposition after the membership function shape change. Finally max composition,
The center of gravity is calculated and the final output is obtained.

The first invention will be described by taking the air pollution prediction system for an automobile tunnel as a specific example. The air pollution prediction system for an automobile tunnel is a system for predicting and outputting the air pollution degree after 5 minutes by inputting the value of the current state variable in the tunnel in a one-way automobile tunnel. The actual input is the number of cars currently running in the tunnel, the running speed, the change in running speed, the speed of the natural wind at the tunnel entrance, etc., and the output is the soot transmission rate (optical of the atmosphere after 5 minutes). Transmittance). The air pollution in the tunnel is caused by the exhaust gas of automobiles, and the pollution degree increases as the number of automobiles increases. Further, as the speed of the automobile increases, the degree of pollution decreases due to the ventilation effect (piston effect) due to the wind pressure of the traveling automobile. FIG. 3 shows a part of the fuzzy rule created based on these phenomena. Moreover, the fuzzy rule of this air pollution prediction system is called a basic rule after this.

Next, regarding the inference control mechanism, a fuzzy inference system is used for the inference control mechanism. The input of the inference control mechanism is the goodness of fit of the antecedent part proposition and the suitability of the rule antecedent part, and the output is the shape parameter change amount and rule weight of each membership function. The fuzzy rule that changes the membership function in the inference control mechanism has the effect of correcting the shape of the membership function that changes according to the input, that is, the control law that changes according to the state in the tunnel. For example, it has been known that when the natural wind blows in the direction opposite to the traveling direction of the automobile, the ventilation effect due to the wind pressure of the automobile decreases as compared with the case where the natural wind blows in the same direction. This means that the antecedent part propositions appearing in the rule in FIG. 3, “the vehicle speed is fast” and “the vehicle speed is slow”, change as shown in FIG. 4, respectively.

Further, the rule for changing the rule weight in the inference control mechanism has the effect of suppressing the contradiction between rules in the inference process and improving the accuracy of inference. For example, for a certain input, rule 10 and rule 11 may be established at the same time. This is because rule 10 and rule 11 each have two propositions in the antecedent part, but the consequent parts of these two rules are inconsistent with each other. Therefore, in order to prevent the two rules from being established at the same time, the inference control mechanism uses the "if
Rule 8 is established then Rule 10 is reduced in weight and
The weight of rule 11 increases ”,“ if rule 8 fails
Then, prepare fuzzy rules such as "weight of rule 10 increases and weight of rule 11 decreases". By preparing this rule, inference can be performed without contradiction. An example of the fuzzy rule of the inference control mechanism is shown in FIG.

There are 74 cars now and the speed is 92km /
s, and the wind speed of natural wind is -5 m / s. At this time, the suitability of the antecedent part proposition "The wind direction of the natural wind is opposite" of Rule 9 becomes 0.7 as shown in FIG. As a result, the rule of the inference control mechanism "if the natural wind direction is reversed and then the vehicle speed membership function shifts to the right."
Is satisfied with. As a result, the membership function of "the vehicle speed is fast" in the antecedent of the rule 2 is corrected as shown in FIG.
As a result of the recalculation, the goodness of fit of rule 2 changes from 1.0 to 0.7. Further, in the inference control mechanism, the rule weight is calculated from the rule antecedent conformity, but the rule weight is changed in the direction in which the soot transmittance decreases as in the above example. As a result of these actions, the inference output is corrected as shown in FIG.

In the above embodiment, the input of the inference control mechanism is the antecedent part proposition conformity, the rule antecedent part conformity, and the output is the parameter change amount of the membership function and the rule weight.
In addition to this, inputs to the inference control mechanism can be inputs to the system, rule conformance, outputs, and the like. Further, the output of the inference control mechanism includes a scaling parameter of the proposition, a weight of the proposition, and the like. Of course, it is also possible to combine a plurality of these.

In the first embodiment described above, fuzzy inference is used as the inference control mechanism, but it is of course possible to use a neural network or a production system instead of fuzzy inference.

The second invention will be described. The dotted line in FIG. 9 is a block diagram of a fuzzy inference apparatus having an inference control mechanism similar to that of the first embodiment. The inference control mechanism is a simple perceptron (hereinafter,
Simply called a perceptron). As shown in FIG. 10, the perceptron has a structure in which an input 1001 is coupled to an output neuron 1002 via a weighted link 1003. The output neuron has an input / output characteristic as shown in FIG. 11, and outputs a value of 0 to 1 according to the weighted sum of inputs. The perceptron has a pattern discrimination ability to classify an input signal into two clusters by supervised learning. Details of the ability and a learning method are described in, for example, Hideki Aso, "Neural Network and Information Processing". There is.

The inference control mechanism uses the pattern discrimination ability of the perceptron to judge whether each rule is valid or invalid for the current input state from the pattern of the antecedent part proposition fitness. Weight rules. As shown in FIG. 12, the inference control mechanism has a form in which the number of inputs is the number of antecedent propositions and the number of perceptrons having one output is the same as the number of rules. The input of each perceptron is the fitness of the antecedent proposition, and the output is the fitness of each rule obtained from the pattern of the antecedent proposition fitness.

Learning of the perceptron, which is an inference control mechanism, is performed by combining the extended backpropagation method in the fuzzy inference system and the correlation learning method of the perceptron. In the extended backpropagation method, in the fuzzy reasoning, the error of each inference output from the error of the inference output
The error of each proposition goodness of fit and the error are calculated in the reverse order of the reasoning
Finally, it is a method of changing the shape of the membership function from the error of the propositional fitness. Details of the extended backpropagation method are described in IPSJ Microcomputer and Workstation Workshop Proceedings (91-MIC-66-5). If this extended backpropagation method is used, the error of each rule conformance can be calculated from the error of the inference output.
Correlation learning of the perceptron is performed using the error of each rule conformity as a teacher signal. In correlation learning, the weight of each link is changed in proportion to the product of the input of the link and the teacher signal.
In FIG. 12, the change amount Δwij of the weight of the link 1203 that connects the input from the proposition i of 1202 and the neuron that outputs the weight of the rule j of 1204 is as follows.

[0030]

[Mathematical formula-see original document] Δwij = η · Oi · e where Oi is the output of proposition i, e is the error of the antecedent fitness of rule j, and η is a learning coefficient. As a result, the link weight wij after the (k + 1) th learning is given by the following equation.

[0031]

## EQU00002 ## wij (k + 1) = wij (k) +. DELTA.wij (k) Perceptron learning is performed by preparing an appropriate number of input / output groups. In the one-time learning, the weight of each link is changed by a small amount, and the learning is stopped when the error between the output of the perceptron and the teacher signal converges.

A learning method for the entire system is shown in FIG. In learning, first, a normal fuzzy inference or a fuzzy inference according to the first embodiment or the like is performed to obtain an inference output. Next, the error in the rule weight 906 is calculated by the extended backpropagation method. The inference control mechanism 910 uses this error as a teacher signal.
Learn. In the inference control mechanism 910, the link weight of the perceptron is changed according to the equations (1) and (2). The error of the inference output at the time of the previous learning is stored, and when the change in the error becomes sufficiently small, it is considered that the learning has converged and the processing is terminated.

14 and 15 show the results of learning performed on the same rule as that of the first embodiment of the present invention. FIG. 14 shows the change in the inference output error due to learning, and FIG. 15 shows the inference output before and after learning. It can be seen from FIG. 14 that learning can reduce the error in the inference output, and from FIG. 15 that accurate inference can be achieved by learning. Fig. 16 shows an example of the rule antecedent conformance, the perceptron output, and the effective rule antecedent conformance in the inference process.
It can be seen from Fig. 16 that the perceptron switches the rules according to the pattern of the goodness of fit of the antecedent part proposition.

In the above embodiment, the simple perceptron is used as the neural network of the inference control mechanism unit. However, a neural network such as a multi-layer perceptron, Boltzmann machine, Hopfield type model, etc. can be used as well. In addition to the above correlation learning, it is possible to use a learning method such as orthogonal learning, LVQ, back propagation, competitive learning, or the like.

Further, in the above embodiment, the input of the neural network is the antecedent part proposition conformity, the output is the rule weight, and the teacher signal is the antecedent part conformance error. Condition part suitability, rule suitability, inference output, etc., one or more of a system input value scale value, membership function shape, etc. as output, and inference output error, proposition suitability error as a teacher signal, It is possible to use one or more of the error converted into the inference system input value.

Further, although the neural network is used as the inference control mechanism in the above embodiment, a fuzzy inference system, a production system or the like is used as the inference control mechanism.
It is also possible to use the learning method such as the extended backpropagation described above for learning.

A third aspect of the present invention is to generate a new rule by analyzing the weight of each link after learning when a neural net is used as the inference control mechanism in the second aspect of the invention. As an example, the third invention will be described using a fuzzy reasoning apparatus having the same structure as the description of the second invention. That is, a simple perceptron is used as the inference control mechanism, the antecedent part proposition fitness is used as the input of the inference control mechanism, and the output of the inference control mechanism is the weight of each rule. In learning, the error with respect to the correct answer value of the rule weight is obtained by the extended backpropagation method, and perceptron correlation learning is performed using this as a teacher signal.

If learning is performed under the above conditions,
From Equation 2, the following is clear about the weight of each link.

1 When the weight of the link wij connecting from the antecedent part proposition i to the rule j is positive and large, the antecedent part proposition i and the consequent part of the rule j have a strong positive correlation. That is, when the rule j is satisfied, the suitability of the antecedent part proposition i is high.

2 When the weight of the link wij connecting the antecedent part proposition i to the rule j is negative and large, the antecedent part proposition i and the consequent part of the rule j have a strong negative correlation. That is, if rule j is satisfied, then rule j
Does not hold.

FIG. 17 shows a part of the weight of the link after learning in the second invention. This is the weight of the connection link of the rule "If the wind speed change rate of natural wind decreases then the smoke transmission rate increases", but the antecedent part proposition "The change in the number of small cars is 0" is combined with a positive weight, The section proposition "Natural wind speed is 0" is combined with a negative weight. Above 1,
From 2, this rule can be rewritten to include the three antecedent propositions. The rewritten rule is "if the wind speed of natural wind is not zero, the rate of change in wind speed of natural wind is reduced, and the change in the number of small cars is 0 then the soot transmission rate is increased."

As described above, according to the present invention, even when there is only piecemeal knowledge about the inference target, a set of input / output is prepared for learning and a rule is generated by analyzing the learning result. You can In particular, it can be used to acquire the tacit knowledge possessed by a skilled person, which has been a problem until now, which cannot be easily grasped by interviewing a skilled person.

In the above-described third embodiment of the present invention, the correlation between the antecedent part proposition conformity and each rule is examined by using the input of the inference control mechanism as the antecedent part proposition conformity and the output of the inference control mechanism as the weight of the rule. However, it is also possible to examine the correlation of the goodness of fit between the rules by using the input of the inference control mechanism as the antecedent degree of the rule and the output of the inference control mechanism as the weight of the rule. In this case, the degree of similarity and the degree of contradiction of each rule can be known by examining the learning result.

Further, in the above embodiment, the simple perceptron is used for the neural network of the inference control mechanism, but other than this, it is possible to use a neural network such as a multilayer perceptron, Boltzmann machine, Hopfield type model, In addition to the above correlation learning, it is possible to use a learning method such as orthogonal learning, LVQ, back propagation, competitive learning, or the like as the learning method. When these neural nets and learning methods are used, not only the weight of each link is simply analyzed, but also dynamic analysis such as analysis of the firing state of each neuron of the neural network may be performed together.

[0045]

As described above, according to the present invention, the inference control mechanism is provided in the fuzzy inference, the state of the inference target is determined, and the inference is performed using the shape, rule, etc. of the membership function most suitable for the determined state. Since it is performed, it has the effect of enabling precise inference even when the inference target has a complicated structure.

Further, according to the present invention, since the state of the inference target can be automatically determined by learning, there is an effect that precise fuzzy inference can be performed even when there is only a piece of knowledge about the inference target. .. Furthermore, there is an effect that the hidden knowledge can be acquired by specifying the shape of the membership function, the rule, etc. for each state of the inference target by learning.

[Brief description of drawings]

FIG. 1 is an example of a block configuration diagram of a fuzzy inference apparatus having an inference control mechanism of a first invention.

FIG. 2 is a reasoning procedure in the fuzzy reasoning apparatus of FIG.

FIG. 3 is a part of a fuzzy rule of the system used in the description of the first embodiment.

FIG. 4 is an example of correcting the membership function when the state of the inference target changes.

FIG. 5 is a part of a fuzzy rule of an inference control mechanism for correcting a membership function when a state of an inference target changes.

FIG. 6 is an example of input of fuzzy inference in an inference control mechanism.

FIG. 7 is an example of correcting a membership function by an inference control mechanism.

FIG. 8 shows the effect of correction by the inference control mechanism on the inference output.

FIG. 9 is an example of a block configuration diagram of a fuzzy inference system having an inference control mechanism having a learning function of the second invention.

FIG. 10 is a principle diagram of a perceptron used for an inference control mechanism in the second embodiment.

FIG. 11 is an input / output characteristic of a neuron used in the perceptron of the second embodiment.

FIG. 12 is a perceptron of an inference control mechanism in the second embodiment.

FIG. 13 is a procedure for learning an inference control mechanism.

FIG. 14 is a diagram showing a change in error due to learning when actually learning is performed in the second embodiment.

FIG. 15 is a diagram showing improvement of inference accuracy by learning in the case of an embodiment.

FIG. 16 is a diagram showing switching of rules by the inference control mechanism after learning.

FIG. 17 shows a part of the link weights of the perceptron of the inference control mechanism after learning.

[Explanation of symbols]

1201 ... Input from the antecedent part proposition fitness, 1202 ... Perceptron input, 1203 ... Weighted link, 1204 ... Output neuron, 1205 Rule weight output.

Front Page Continuation (72) Inventor Toshihiko Nakano 5-2-1 Omika-cho, Hitachi-shi, Ibaraki Hitachi Ltd. Omika Plant

Claims (9)

[Claims] 1. A storage device for storing the shapes of membership functions of propositions of the antecedent part and consequent part of each rule, and a propositional fitness calculation device for calculating the fitness of the proposition of each rule from an input. An antecedent part fitness calculation device for calculating the suitability of the antecedent part for each of the rules, and at least a part of each of the adaptability obtained in the fuzzy inference process as an input, and an inference parameter of the fuzzy inference according to the input And a reasoning control mechanism for changing at least a part of the reasoning, and fuzzy reasoning is performed again using the changed reasoning parameters. 2. The inference parameter to be changed is at least one of a scale value of a proposition, a shape parameter of a membership function, a weight of a proposition, and a weight of a rule. Fuzzy reasoning controller. 3. The fuzzy inference apparatus according to claim 1 or 2, wherein the input of the inference control mechanism is the input and output of the fuzzy inference apparatus, the degree of proposition matching in the fuzzy reasoning process, and the degree of matching of the antecedent part of a rule. , A fuzzy inference control method characterized by using one or more of the conformity of rules. 4. The fuzzy inference apparatus according to claim 1, wherein fuzzy inference is performed using a plurality of fuzzy inference control methods. 5. The fuzzy inference apparatus according to claim 1, wherein at least one of a neural network, a production system and a fuzzy inference system is used as an inference control mechanism. .. 6. If the inference result by the fuzzy inference is not desirable, the internal structure of the inference control mechanism controlling the inference associated with the fuzzy inference mechanism is modified so that the desired result is obtained. A learning method for a fuzzy reasoning device. 7. A means for calculating an error between an inference result and a desired output in each step of inference according to claim 6,
A learning method for a fuzzy reasoning apparatus, wherein at least part of the output error and the error in the inference process is used as a teacher signal to learn the inference control mechanism. 8. In the learning of the inference control mechanism according to claim 6 or 7, a knowledge acquisition means for extracting hidden knowledge by analyzing the learned inference control mechanism by a predetermined means is provided. Knowledge acquisition method characterized by that. 9. An automatic rule acquisition function according to claim 8, wherein the extracted knowledge is expressed by a proposition, a rule, or a combination of both.