JPH04205163A - Neural network, its remembering or learning method, and fuzzy inference device using the same - Google Patents

Abstract

PURPOSE:To obtain the fuzzy inference device which is applicable to the prediction, diagnosis, control, etc., of a process and further stable and highly accurate by providing a fuzzy qualitative evaluating means which evaluates input information qualitatively and supplying >=1 qualitative evaluated value to an input layer for each piece of input information. CONSTITUTION:A 1st fuzzy qualitative evaluating means 2 after evaluating the input information qualitatively with a membership function determined by experimental knowledge and normalizing a dynamic range inputs the result to the input layer of the neural network 3. A 2nd fuzzy qualitative evaluating means 4 performs a quantizing process by using the output value of the neural network 3 as the qualitative evaluated value of the experimentally determined membership function. Consequently, even input information having large nonlinearity or much indeterminateness can accurately be evaluated. Further, the information can be normalized to a specific range, so even input information having different property makes learning and remembering stable and highly accurate.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to neural networks, and particularly to a neural network learning and recall method suitable for objects with large uncertainties or strong nonlinearity.

[Conventional technology]

Neural networks are used in information processing systems.
In particular, it enables pattern recognition processing of images, sounds, etc. by learning patterns without using complex algorithms. When training a neural network, input the pattern you want to learn, modify the connections of the neural network to reduce the error between the neural network's recall output and the teacher signal, and repeat this process so that the neural network performs the specified action. It is intended to do so.

The structure and operation of this neural network are explained, for example, in ``Pattern Recognition and Learning Algorithms'' (written by Yoshinori Uesaka and Kazuhiko Ozeki; published by Bun-Sogo Publishing).

One of the important challenges in applying neural networks to real devices is learning and recalling various types of input signals in response to their characteristics, and learning similar or homogeneous input patterns that are different from the learned patterns. The goal is to be adaptable even to other people.

For example, even if speech is the same word pattern, its duration differs each time it occurs or depending on the speaker.
You will need to be creative with your input. As a countermeasure to this problem, JP-A-1-2
Publication No. 41667 discloses that by dynamically normalizing the time axis, fluctuations in the generation time length of unknown audio data are
It describes training a neural network that can recognize the same words as the same.

Further, Japanese Patent Laid-Open No. 2-98770 describes a learning method for a multilayer neural network that specifically responds to input patterns by nonlinearly processing the input patterns using a Gaussian filter.

[Problem to be solved by the invention]

The above-mentioned conventional techniques often provide good results when the input information is of the same quality, such as audio or image signals, or has a strong correlation. However, when trying to apply neural networks to control or diagnosis of processes where various types of process variables are input, oscillations occur during learning and stable convergence is often not possible.
There were difficulties in putting it into practical use.

For example, when predicting pollution inside a road tunnel, completely different information such as the number of vehicles traveling and natural wind are part of the input. The former is 0 to 200 in units of 15 minutes.
The latter has a range of -5 to +5 in m/s.
It has a range of The influence of the former on the pollution value is extremely non-linear, with almost no effect from 0 to 10 [15 minutes for cars], the degree of influence becomes significant and approaches linear from 30 to 70 [15 minutes for cars], and from 70 (15 minutes for cars) ] Above this, the speed slows down again.Moreover, the nonlinearity becomes extreme when a large truck or the like that releases a large amount of smoke is running.

On the other hand, the latter has a relatively negative influence on the contamination value and has a narrow range.

In this way, when the input information has various dynamic ranges and large uncertainties and nonlinear elements, if multiple pieces of input information are fed directly to the neural network, it will be difficult to extract the features of each piece of information, making it difficult to learn. Does not converge. Moreover, as in the prior art described above, there are only a limited number of objects for which input information can be subjected to nonlinear processing using a specific filter, and there are often no clear means for extracting features.

Furthermore, in processes such as process control, it is required to convert the recall output of a neural network from a crisp value to a continuous quantity, and it is difficult to quantify when there are strong nonlinear elements.

An object of the present invention is to overcome the above-mentioned problems and provide a stable and highly accurate neural network and its learning method, which can be applied to process prediction, diagnosis, control, etc.

[Means to solve the problem]

In order to achieve the above object, the present invention is characterized by being constructed by the following means that introduces the fuzzy set concept.

That is, the present invention includes an input layer for inputting information, an output layer for outputting the final result, one or more intermediate layers between the two layers,
In a neural network having a plurality of neurons in each layer and synapses having weighting coefficients that connect neurons in adjacent layers, the input information is qualitatively evaluated using a fuzzy membership function determined in advance by a priori knowledge. Means for inputting the dynamic range to the input layer after normalization is provided.

Furthermore, means is provided for performing a quantification process by using the output value of the neural network as a qualitative evaluation value of a fuzzy membership function determined a priori.

[Effect]

In the present invention, processing within a neural network is performed not by the absolute value of input information as in a linear model, but by extracting its features, and each input value of the input layer neuron is of the same dimension (unit: It is based on the idea that there is no need to treat it in terms of dimensions.

According to the configuration of the present invention described above, even when input information has various ranges and includes strong nonlinearity and many uncertain elements, the correlation between the input value and the recall result can be understood to some extent and qualitative expression can be performed. When possible, learning/recall is made possible by using qualitative evaluation values evaluated using fuzzy membership functions, which are qualitative expressions, as input information to the neural network.

In addition, even if the recall result of a neural network contains strong nonlinear elements, the recall result shows a qualitative evaluation value, so if the correlation between the evaluation value and quantitative value is known to some extent, the output value of this neural network It becomes possible to obtain a quantitative value output by defuzzing using as an evaluation value.

As described above, according to the present invention, the present invention can be applied to various types of information such as process diagnosis and control, which were previously thought to be impossible to apply, by applying the present invention according to the nature of each information. It becomes possible to learn/recall neural networks that respond correctly to various characteristics of objects. Furthermore, since nonlinear processing is processed using a fuzzy set concept, it is possible if the correlation between input information and recall results is known to some extent from experience or experimentation, and has great applicability.

Furthermore, by using a neural network configuration that combines the above-described fuzzy evaluation of input information and quantification of recall results, the inference rules in conventional fuzzy inference means can be replaced by this neural network. This makes it possible to shorten the comparison tuning time and increase accuracy in conventional fuzzy inference means.

〔Example〕

Hereinafter, preferred embodiments of the present invention will be described based on the drawings. 1 to 10 illustrate a first embodiment of the present invention, and FIG. 1 shows the overall configuration of a neural network learning device. The learning device 1 includes a first fuzzy qualitative evaluation means 2 for qualitatively evaluating the learning pattern X, a forward neural network 3, a second fuzzy qualitative evaluation means 4 for qualitatively evaluating the teacher pattern d(X), and an error backpropagation weight. It consists of a correction means 5 and a computer-based control means 6 for managing and controlling the operation of each of these means.

The configuration of the first fuzzy qualitative evaluation means 2 in FIG. 2 is shown,
Input information of means 2 Xa = (x□, X2゜..., xn)
, and outputs n times the number of qualitative evaluation values.

FIG. 3 conceptually shows the configuration of the forward neural network 3. Although a forward-facing three-layer model is illustrated, the present invention is equally applicable to a multi-layer structure of four or more layers.

Neural net 3 has n input layer neurons x (1
,2,...tn), 1m hidden layer neurons y (j=1,2...+m), and 0 output layer neurons z (k=1,2...+m).・, Q) and a synapse having a weight WIJ that connects the input layer and the hidden layer,
It consists of a synapse having a weight Vth that connects the intermediate layer and the output layer. Weight wlj, Vjh (7) 17
'/ji is -1.0 to 1.01? be. Each neuron has spatial additivity and nonlinearity, and performs processing (recall) that will be described later.

FIG. 4 is a diagram illustrating the operation of the goodness-of-fit calculation means 21. As shown, the means 21 includes, for example.

Membership functions corresponding to three types of qualitative evaluations such as S (small), 2M (medium), and B (large) are stored in advance from a priori information obtained from experience, experiments, etc. Input information XL (i=1.2...tn) is evaluated using three types of curves, μ5(Xi)”Xts (μS is the number of fitness degrees for evaluation type S) μM(xt) = XIM μa (xt) =XIB and converted into three evaluation values, and these qualitative evaluation value set X
' becomes the input to the forward neural network 3.

On the other hand, the second fuzzy qualitative evaluation means 4 converts the teacher pattern d (X) into three types of evaluation values for each piece of information in d (X), resulting in a qualitative evaluation value set d' (X).

The error backpropagation correction means 5 includes the fuzzy qualitative evaluation means 2,
4 and the recall output of neural network 3 h (X
) as input, the neural network 3
Modify the synaps weights of .

The learning device configured as described above has a learning pattern evaluation value set X''' (X1S+XxHg xIB,
X2StX2M, xzst...) and a set of evaluation values d' (X) for the teacher pattern are given, and learning is performed as follows.

■ Intermediate layer neuron output value calculation After calculating the input layer neuron output value, that is, the product of each input value of the evaluation value set of the learning pattern and the synapse weight WsJ, calculate their sum, and calculate the nonlinear function (
Here, the sigmoid function) is calculated to obtain the output value of the hidden layer neuron.

'/J=σ(woJ+wlJsx1s+w, JMχ, H
+W, JBX, B+w, Jsx2s+W2aHXzM+
...+W□s x na) j=1.2. -...9m Where, y: Output value of the hidden layer neuron σ: Sigmoid function σ(S)''1/ (1+e-S) ■ Output layer neuron output value calculation The hidden layer neuron output value and weight vJk obtained above After calculating the products of , calculate the sum of them, and further calculate a nonlinear function (sigmoid function) for the sum to obtain the output layer neuron output value.

zk:σ(vok+v1kyl×2ky2+゛.

10v, 1ymaro) k=1. 2. ..., Q The set of output values Zb of the neural network 3 obtained by the above processing, that is, the recall pattern h (X) is the first
The error backpropagation weight correction means 2 shown in the figure performs the following processing to correct the weights, and by repeating this process, the neural network is trained.

■ Synapse weight correction amount calculation Synapse weight correction amounts δ, k(x) and δ□ that minimize the error between the evaluation value d' (X) of the known teacher pattern and h (x) obtained in the above ■ , (X).

62k (x) = a (x) (h h (x) d k
(x)) ・b h(x) (1-d h(x)) f J(x) (1-f J(x)) (j=L,...,m) Here, α(x ) is the output vector h for the input vector X'
This is the degree of importance related to the proximity of (x) and is a known item.

fa(x) is a function of m n variables, fJ(x)=δ(w, B+ W, J #x1+-+Wn
J'
Using the results of , we create new weights v, n e W.

Correct to “WIJn6”.

8F = Oy..., m; = 1. ..., Q) Δ (i = o, ..., n; j = 1. ... + m) where,
IXI is the number of learning pattern elements.

Δ is the differential width and is a known item.

Since the learning patterns and teacher patterns that are learned using the above-mentioned learning device are qualitatively evaluated and normalized by a membership function of 0 to 1, the degree of correlation consistency between the learning patterns and the teacher patterns is improved. This enables stable and highly accurate learning.

Further, even with a learning device using only one of the first qualitative evaluation means or the second qualitative evaluation means 4, similar effects can be obtained depending on the object to be recalled.

Figure 5 shows an example in which qualitative evaluation is performed only on input;
shows the configuration of the learning device.

Figure (b) explains the operation in this configuration.
An example in which learning to predict changes in indoor temperature by an air conditioner is performed using known indoor temperature displacement ΔTin, outdoor temperature displacement ΔTout, and cooler operating displacement ΔCL as learning information, and known indoor temperature change value ΔT1- after a certain period of time as teacher information. It is. In this case, as shown in the same figure (c), the temperature variation is ±5
, the displacement of the cooler is in the range of ±20, and each has different nonlinear characteristics. Therefore, if each piece of information is input as is, it may be difficult to converge.

Therefore, each input is qualitatively evaluated (also normalized) using a predetermined fuzzy membership function based on experience and actual measured values, and the characteristics of each information are well extracted by inputting the evaluation values into a neural network. learning accuracy,
Both efficiency can be improved.

Figure 6 is an example of qualitative evaluation of only the output;
indicates the configuration of the learning device. This configuration is suitable for time-series data with homogeneous information (sound, one image, etc.). Figure (b) shows a data group (variation range Δ
P, -n; n=o,i.

2,...) to predict the next day's stock price fluctuations. If the form of change in the past is oscillating or gradually expanding (triangular change), it is a nonlinear situation change, and if the output is defined as a recall value of 1 new ron, convergence will be slow and accuracy will be poor.

Therefore, the number of output neurons is set to three, for example, and evaluation types of increase (+), unchanged (O), and decrease (-) are defined for each. On the other hand, the next day's known stock price fluctuation range ΔPi which is teacher information
Qualitatively evaluate es using fuzzy qualitative evaluation means 4 and perform the above 3
Obtain the evaluation value of each type and use this as the teacher pattern. This allows the characteristics of nonlinearity between input data to be well captured, and
Accuracy of learning.

Both speed improves. This is because the nonlinear grasping power originally possessed by forward-looking neural networks can be easily exploited by defining the output neurons as qualitative values.

After completing learning using the above steps, the synapse weights are fixed and the neural network is transplanted to the actual device. The operation in an actual device is called recollection, and it allows accurate judgment to be made even in response to unknown input information. The operation is the same as in steps ① and ② above. That is, after calculating the sum of the products of the input information and synapse weights, the intermediate layer neuron output value is determined by calculating the sigmoid function value. Next, in a similar manner, output layer neuron values, that is, recall results, are obtained using the intermediate layer neuron output values and synapse weights.

FIG. 7 shows the configuration of the recollection device 10 according to the present invention. The recall device 110 inputs information X'' (Xl,
2. -, XI, , -XN) as input and outputs a qualitative evaluation set.
A neural network 3 receives the qualitative evaluation set X' outputted from , and outputs the qualitative evaluation value set h(x) for the recall target, and the neural network 3 receives the qualitative evaluation value set h(x) as input and converts it into a quantitative value. It is composed of a quantification means 6 that outputs the following information. Of these, the means 2 and the neural network 3 perform the same operations as the learning device N1 shown in FIG. 1, so their explanation will be omitted.

FIG. 8 explains the operation of the quantification means 6.

In this example, the output layer neuron values of forward neural network 3, that is, the qualitative evaluation of the output values, are: NB (large negative) NM (medium negative) NS (small negative) Z○ (zero) PS (small positive) ) Define seven types of evaluation: PM (medium positive) and PB (high positive). The vertical axis shows the goodness of fit O0O~1,
0, and the horizontal axis is the recall target value H (-HMAX ~ +HM^X
), and the seven types of qualitative evaluation curves are created based on a priori information. The number of output neurons of the forward neural network 3 is set to 7 in this example, and the qualitative evaluations NB and NM are performed.

NS, 20. Make it compatible with PS, PM, and PB. According to the recall result of the neural network 3, the corresponding evaluation curve of the membership function is cut by the neuron output value, and the center of gravity of the area indicated by diagonal lines is determined.

The value of H corresponding to this center of gravity is output.

When recall is performed using such a recall device 6, the input information is qualitatively evaluated and normalized and input to the neural network, and the recall output of the neural network can be treated as a qualitative evaluation value. This recall output can be defuzzed and quantified using a predetermined membership function.

Note that if the recall output is obtained as a quantitative value (as in the example shown in FIG. 5(b) above), or if quantitative conversion is not required, the configuration does not include the quantification means 6 as shown in FIG. 9.

Furthermore, when recollecting a qualitative evaluation value from input data that is not subjected to qualitative evaluation (the example shown in FIG. 6(b) described above), recollection is performed in a configuration without the qualitative evaluation means 2 as shown in FIG.

11 to 18 illustrate an embodiment in which the present invention is applied to an actual process.

FIG. 11 shows the target process, a tunnel ventilation process. The ventilation process in road tunnels has many uncertain elements and large nonlinear elements, making it difficult to obtain good results using conventional control methods.

An example in which the neural network of the present invention is applied to contamination prediction in this process will be described below.

FIG. 12 shows input signals of the contamination prediction device 30.

Contamination transition after a certain period of time (In this example, the VI value, which indicates good visibility, or transparency, is used as a contamination index. The VI value is O
-100 (%), and the closer it gets to 100 (%), the better the visibility) is determined by the displacement of the number of large vehicles (ΔTB, Δ is a symbol indicating displacement, Defined as (current value) - (previous value)), vehicle speed (T
S).

Contamination value displacement (ΔvB), mechanical ventilation force displacement (ΔM).

The main factors include traffic volume displacement (ΔTR) and natural wind (WN). As the latest prediction method, a method based on fuzzy multi-stage inference has been put into practical use, but its accuracy is about 0.72 with a correlation coefficient, as shown in FIG. What is the correlation coefficient?
As shown in Fig. 4, the predicted values VIc+, pre+tcte- and the actual values vIC+, &
This is an index indicating the accuracy of ctu&1.

FIG. 15 shows a contamination value transition learning device according to the present invention.

Learning pattern X is the above-mentioned ΔTB, TS.

The element is X& composed of ΔVIc, ΔM, ΔTR, and WN. Three types of qualitative evaluations are made for this person type information: decrease (-), maintain the status quo (0), and increase (+).
Qualitative evaluation is performed by fuzzy qualitative evaluation means 2 using a membership function determined a priori. Therefore, the number of neurons in the input layer of the neural network 3 is 18. On the other hand, the known contamination value displacement actual value Δy 1 c+, actuat after a certain period of time is also qualitatively evaluated by the fuzzy qualitative evaluation means 4 using a membership function determined a priori.

The error backpropagation weight correcting means 5 uses the information qualitatively evaluated by the means 2 as a learning pattern, inputs the qualitative evaluation value from the means 4 as a teacher pattern, performs learning according to the above-mentioned procedure, and constructs a neural network. Correct the weight of synapses in 3.

The weights of the synapses that have completed learning as described above are fixed, and the neural network is transferred to the actual device.

FIG. 16 is a contamination value prediction (recall) device according to the present invention. The first fuzzy qualitative evaluation means 2 and the forward neural network 3 perform the same operation as the learning device shown in FIG. Forward neural network 3 performs seven types of evaluation NNB for pollution value displacement ΔVI (ΔVIc+□)
(significantly decreased), NNM (ΔVIc+1) (decreased),
NNS (ΔVIc+,) (slightly decreased) v Nzo
(ΔVIc+j (maintaining the status quo).

Np5 (ΔV I Cat ) (Slight increase L NPM
(ΔVIc+, ) (increase), Npa (ΔVIc+t
) (considerably increased) are the outputs of the output layer neurons, respectively. The quantification means 6 takes these as input, uses a membership function determined based on a priori information, and after quantitative conversion, outputs the predicted contamination displacement value ΔV after a certain period of time as the output of the prediction device. 41PrediCLed is output.

Figure 17 shows a specific example of its operation, NNB (AVIc+,) = 0.082NNM (ΔV
Ice,) =0.102NNS(ΔVIc+,)
=0.113 NzoCΔVIc+, )=0.932 Nps(ΔVIc+,)=0.221 NPM(ΔVIc+,) =0.110NPB (Δ
The qualitative evaluation value of VIc+, )=0.085, that is, the output layer neuron output value, gives the quantitative value +1.6 (%)
is predicted to increase.

According to the contamination prediction device of the present invention, as shown in FIG. 18, the correlation coefficient between the predicted value and the measured value is 0.85, which is greatly improved compared to the conventional method.

The contamination prediction device of the present invention described above is one in which the inference rules in the conventional fuzzy inference device are replaced with a neural network, and the recall device of the present invention, that is, the qualitative evaluation of input information and the quantification of the recall result. The neural network of the present invention, which combines the following, realizes a new fuzzy inference means.

〔Effect of the invention〕

The present invention is configured as described above, and input information is qualitatively evaluated using fuzzy sets and used as input to a neural network, so input information that includes strong nonlinearity and a lot of uncertainty can be evaluated accurately. . Moreover, since it is normalized to a given range.

This has the effect of making learning/recall more stable and highly accurate even when input information has different properties.

In addition, the output value of the neural network is treated as a qualitative evaluation value of the target, and the output value is obtained by quantifying it using a membership function determined a priori, making learning/recall more stable/highly accurate. effective.

Furthermore, by combining the above, it becomes possible to replace the inference rules in conventional fuzzy inference means with neural network recollection means, resulting in shorter tuning time and higher accuracy compared to conventional fuzzy systems. It can be achieved.

Furthermore, the present invention enables neural network information processing, which could only be applied to extremely limited targets in the past.
It is possible to achieve the effect of making it applicable as a means of solving a wide range of general problems.

[Brief explanation of the drawing]

FIGS. 1 to 9 explain the first embodiment of the present invention. FIG. 1 is a configuration diagram of the forehead frost value of the neural network, and FIG. 2 is a configuration diagram of the first fuzzy qualitative evaluation means. ,
FIG. 3 is a conceptual illustration of a forward neural network, FIG. 4 is an illustration of membership functions and operations of the fitness calculating means, and FIGS. 5 and 6 are modifications of the learning device of the present invention. FIG. 7 is a configuration diagram of the recollection device of the present invention, FIG. 8 is a diagram explaining the operation of the quantification means,
FIG. 9 and FIG. 10 are configuration diagrams showing modified examples of the recollection device of the present invention, respectively. Figures 11 to 18 are for explaining the second embodiment of the present invention, where Figure 11 is a model diagram of the tunnel ventilation process, and Figure 12 is a diagram explaining input and output signals of the pollution prediction device. , FIG. 13 and FIG. 14 are block diagrams of a learning device of a conventional contamination prediction device, and FIGS. 16 and 17 are block diagrams of a contamination prediction device (recall device) of the present invention and explanatory diagrams of its quantification operation. . FIG. 18 is a diagram explaining the present invention in detail. 1. Learning device, 2. First fuzzy qualitative evaluation means, 3.
Neural network, 4. second fuzzy qualitative evaluation means, 51 weight correction means, 6. Quantification means, 10. Recall device, 3o, contamination prediction device. Figure 2 Figure 3 Figure 4 Figure 5 (b) Figure 5 (c) ΔT, n (room temperature change) ΔT, u, (external temperature change) -20020 (%) 〇 2 × × × 8th Figure 17 Figure 18 v1 actual value (%) v1 predicted value (%) Figure 11 Traffic volume TR (TB: large vehicle old song, TS speed) No. 12
Figure 13 ■1 Preliminary value (%) Figure 14 vi value

Claims (12)

[Claims] 1. In a neural network that has an input layer for inputting information, an output layer for outputting a recall result, and one or more intermediate layers between these layers, the neural network recalls and outputs target information based on the input information. A neural network characterized in that a fuzzy qualitative evaluation means for qualitatively evaluating information is provided, and one or more qualitative evaluation values are given to the input layer for each input information. 2. In claim 1, the fuzzy qualitative evaluation means:
A neural network characterized by having a membership function that is predetermined based on a priori information based on experience, experiments, etc. and has an evaluation value of 0.0 to 1.0. 3. In claim 2, the membership function includes a plurality of membership functions obtained by changing evaluation viewpoints for the same input information, and a plurality of qualitative evaluation values for the input information are output from the fuzzy qualitative evaluation means. A neural network characterized by. 4. Input layer neuron group that inputs 1 or more information, 1
a neuron group in an intermediate layer of a stage or higher, a neuron group in an output layer that recalls the recall result using one or more qualitative evaluation value groups, and a synapse that connects the neuron groups in each layer with a predetermined coupling coefficient; A neural network characterized by being provided with a quantification means for converting information to be recalled into quantitative values from a group of qualitative evaluation values of output layer neurons. 5. 5. The neural network according to claim 4, wherein the quantification means defuzzifies the neural network using a predetermined membership function based on a priori information. 6. A neural network that has an input layer for inputting information, an output layer for outputting recall results, and one or more intermediate layers between these two layers, and outputs recall target information (target information) corresponding to the input information. Fuzzy qualitative evaluation means for qualitatively evaluating the input information;
A neural network characterized by comprising a quantification means for converting the recollection target information obtained as a qualitative evaluation value into a quantitative value. 7. The neural network has an input layer that inputs information, an output layer that outputs the recall results, and one or more intermediate layers between these two layers.The neural network has In a neural network learning method for correcting the connection state, known learning information is subjected to fuzzy qualitative evaluation and applied to the input layer, and the error between the output value of the neural network obtained as a result and the known teacher information that is the reference value is calculated. 1. A method for learning a neural network, characterized in that the connection state is modified so that the connection state is less than or equal to a predetermined value. 8. The neural network has an input layer that inputs information, an output layer that outputs the recall results, and one or more intermediate layers between these two layers.The neural network has In a neural network learning device for modifying weighting coefficients of synapses, a fuzzy qualitative evaluation means qualitatively evaluates known learning information and inputs it into the input layer, and a recall output outputted from the output layer and a known reference value. 1. A learning device for a neural network, comprising: a weighting coefficient correcting means for calculating an error between the weighting coefficient and the teacher information and correcting the weighting coefficient so that the error becomes less than or equal to a predetermined value. 9. In claim 8, when the recollection output is indicated as a qualitative evaluation value of the recollection object, a second fuzzy qualitative evaluation means is provided having a membership function determined by qualitatively evaluating the recollection object in advance, and the means A neural network learning device characterized by showing teacher information as a qualitative evaluation value. 10. In a neural network recall method that includes an input layer for inputting information, an output layer for outputting recall results, and one or more intermediate layers between these layers, the neural network outputs a recall target based on the input information. When the information is converted into one or more qualitative evaluation values by a fuzzy qualitative evaluation means, and the output of the neural network based on the qualitative evaluation value indicates the qualitative evaluation value of the recollection object, the output evaluation value is used to convert the information into one or more qualitative evaluation values. A neural network recall method characterized by obtaining a quantitative value of. 11. In a fuzzy inference method that qualitatively infers target information from input information, the weight of each synapse that determines the connection state of a neural network is learned in advance from the known value of the input information and the corresponding known value of the target information. converting each of the input information into a qualitative evaluation value using a fuzzy membership function predetermined for each evaluation type, and inputting each of the converted input information evaluation values to each neuron of the input layer of the neural network; A qualitative evaluation value for each evaluation type of the objective information is obtained from each neuron of the output layer of the neural network by recollection of the neural network based on the input, and a set of the obtained qualitative evaluation values is defuzzed. A fuzzy inference method characterized in that the quantitative value of the objective information is inferred by 12. In a tunnel pollution prediction device that predicts pollution values after a predetermined time from process input information, process quantities such as the known number of vehicles passing, pollution value displacement, ventilation force, etc. are used as learning information, and the known pollution values after a predetermined time are used as learning information. A neural network that has completed training in advance using values as teacher information and a membership function predetermined for each process quantity convert each of the multiple input process quantities into qualitative evaluation values, and the neural network A fuzzy qualitative evaluation means is provided as an input signal of the network, and a quantification means for defuzzifying a set of qualitative evaluation values for contamination values after a predetermined time outputted from the neural network to obtain a quantitative value of the contamination values. Features: Tunnel pollution prediction device.