Vous êtes sur la page 1sur 4

A Design of Genetic Fog Occurrence Forecasting System by Using

LVQ Network
Y. Mitsukura, M. Fukumi and N. Akamatsu
Faculty of Engineering
University of Tokushima

ABSTRACT from 1st layer to 4th layer) in order. Here, the atmo-
A transportation development in recent years is spheric pressure is one based on the international stan-
quite remarkable. However, poor visibility often cause dard atmosphere. The tempreture means the ground
a n accident. Therefore, it i s very important t o fore- temperature. The dew-point means “how much de-
cast a fog occurrence. I n this paper, we propose a grees is vapor saturated in when temperature is low
scheme t o forecast a fog occurrence b3 using the Learn- under an atmospheric pressure ?”. The wind direc-
ing Vector Quantization (LVQ) and a Genetic Algo- tions are defined as follovis:
rithm (GA). This scheme forecasts the fog occurrence The angle of 90: the wind is blowing from the east to
by the weather data which are provided f i o m the Japan the west.
Meteorological Agency. First, the provided data for- The angle of 180: the wind is blowing from the south
mation are shown. Next, the prediction scheme is to the north.
described in detail. I n this method, input attributes The angle of 270: the wind is blowing from the west
f o r a L V Q network are selected by real-coded G A t o to the east.
improve forecast accuracy. firthermore, a partial se- And also, the maximum instantaneous wind velocity
lection processing in the real-coded G A improves its is reported only when an instantaneous wind velocity
convergence properties. Finally, in order t o show the more than 10 knots compared with the average of the
eflectiveness of the proposed prediction scheme, com- wind velocity is observed for ten minutes. Further-
puter simulations are performed. more, the visibility is reported as “9999” when it has
more than 10000.
1 Introduction For example, the data of January 1, 1984, 0:O a.m.
In this paper, we propose a new method to pre- and 0:30 a.m. are shown as follows:
dict a fog occurrence by using the Learning Vector age, month, day, time, atmospheric pressure, tempre-
Quantization (LVQ)[l] - [3] and a Genetic Algorithm ture, dew-point, derection of the wind, wind velocity,
(GA)[4],[5]. The objection in this paper is to predict max instantaneous wind velocity, wind change 1, wind
the fog occurrence precisely by using the LVQ method change 2, visibility, weather, cloud information from
and the GA. 1st layer, cloud information from 2nd layer, cloud in-
The LVQ methods have been utilized in many fields formation from 3rd layer., cloud information from 4th
of engineering researches [1]-[3].However all input at- layer
tributes are used to determine a winner unit. It is not 1984, 1, 1, 0.0, 1009, -2.0, -3.0, 180, 5, -1, -1 -1, 9999,
very good for a high recognition accuracy and a rule 21, 0 2 7, 0 4 25, -1 -1-I., -1 -1 -1, 1
generation from the LVQ network. Therefore in this 1984, 1, 1, 0.5, 1018, -2.0, -3.0, -1, -1, -1, -1 -1, 30, 26,
paper, a method to select the input attributes using 0 2 7, 0 4 25, -1 -1 -1,-1 -1-1, 1
the real coded GA is presented. This method leads where “-1” denotes non-data.
to the rule generation from the LVQ network by pro- Consider the case of the upper data which are mea-
ducing a small-sized network [8]. Relationships among sured at 0 a.m. on January l, 1984. The atmospheric
the input attributes can be defined in a resulting LVQ pressure is 1009 hPa, the tempreture is -2 degrees and
network obtained using the real coded GA. dew-point is -3 degrees. The wind derection is the an-
The data were provided from the Japan Meteoro- gle of 180. This means the wind is blowing from the
logical Agency and there is the one for 12 years( 1984- south to the north. Fur1;hermore “5” means the wind
1995). It consists of the data measured every thirty- blowing at the speed of 5 meters per second. The next
minutes. The data consist of (year) (month) (day) “-1” is the maximum instantaneous wind velocity and
(time) (atmospheric pressure) (tempreture) (dew- is not measured. In the same way, two “-1” means that
point) ( wind direction ) (wind velocity) (maximum the wind change is not observed. The visibility “9999”
instantaneous wind velocity) ( change of the wind di- means that it is more than 10000 meters. The present
rection 1,2) (visibility) (weather) (cloud information weather is “21”, which means that it is raining. Fur-

0-7803-6583-6/001$10.00 0 2000 IEEE


3678
thermore, when there are clouds in different heights,
cloud information of the first layer, the second layer,
the third layer and the fourth layer is displayed in or-
der from the lower layer. The fog occurrence condition
are defined as following:
1. The distance of visibility is 0-1000 meters
2. The weather signal is from 40 to 49.
In order to show the effectiveness of the proposed
I Output layer

scheme, computer simulations are done by using these


data.
2 LVQ network
The object of the LVQ learning proposed by Koho-
nen is to perform pattern classification of input data.
For this input pattern classification, neuron weights
are formed by a supervised competitive learning. The
structure of a LVQ network is shown in Fig.1. Input layer
The LVQ network has an input layer and an out-
put layer. The input and the output layers are fully-
connected by weight vectors. Here, LVQ algorithms
are shown as follows: x1 x2 x3 x4 X”
1. Random values are set as initial weight values. Fig.1 The structure of the LVQ network.
2. An input vector{X = ( z I , z p , ...,.,)}is set in the
3 Inputs selection of the LVQ network
input layer.
by GA
3. The distance between the weight vector of neuron In this paper, inputs to the LVQ network are se-
j and the input vector is calculated as follows: lected by using real cording GA. The procedure of GA
is shown as follows:

Step1 An initial uniform random number ni in [0,1] is


assigned to all genes on chromosomes. These ran-
where, wji denotes the weight connecting the i-th dom numbers (nl,...,n N )are used in the distance
neuron in an input layer to the j-th neuron in the calculation as follows:
output layer.
I N
4. The weight vector which has the smallest d j
wji
is selected, and is renewed according to the fol-
lowing equation.

wji(k) + a(.* - wji)


This is an improved version of the equation (1).

: The result of recognition is true Step2 Calculation of the fitness function { f k , ~ } , where
wji(k+l) = k , 1 denotes the 1-th individual on the k-th gener-
w j i ( k ) - Q ( Z ~- w j i ) ation.

: The result of recognition is false Step3 Individuals are rearranged in turn on the basis of
their recognition rates.
where cr denotes a learning late and k is a time
index. Step4 Take the average of the values of genes at the
same position on chromosomes in the half of in-
This algorithm is called the LVQ1. By the way, in dividuals which has a high recognition rate.
addition to this LVQ1, there are 2 methods called the
LVQ2 and the LVQ3[1] - [3]. As for the details, the Step5 Substitute “0” for the gene which has the small
LVQ2 and the LVQ3 are omitted. In this paper, we ap- value in the average calculated in StepS.
plied the method combining the LVQl with the LVQ3
to the fog occurrence forecasting. Step6 Reproduction of the next generation’s population
through selection, crossover and mutation.

3679
In Stepl, individuals have the same number of at- For the purpose of comparison, we applied the
tributes. In Step2, we use the following fitness func- Back-Propagation (BP) method as a conventional one
tion: to a fully-connected three layered network (NN). The
parameters of NN are shown in the following.
fk,l = A/(the number of individuals)
The number of layers : 3
+B/(the number of individuals) (3)
The number of units in the input layer : 26 x 6
where A is the ranking number which is assigned in (Maximum)
turn by high recognition rate in the LVQ network and
B is the ranking number which is assigned by the num- The number of units in the hidden layer : 20
ber of non “0” in the chromosome of an individual in
GA. In Step3 and Step4, while a generation cycle is The number of units in the output layer : 2
repeated, the high average are assigned to attributes And also, we use a method combining the LVQl with
(chromosomes) which is necessary for the recognition. the LVQ3. The parameters of (LVQl+LVQ3) are
In Step5, the attribute which isn’t necessary for the shown in the following.
recognition can be removed. In Step6, the crossover is
done according to the reference [6],[7]and the muta- The number of units in the input layer : 26 x 6
tion is done as regeneration of the removed attributes. (Maximum)
That is, if the value of a gene is not “O”, then an
uniform random number ni in [0,1]is assigned to it. The number of units in the output layer : 20
3.1 Partial Selection Finally, we apply the proposed scheme to the fog oc-
Average of the chromosome information of the in- currence forecast, in which GA performs a kind of fea-
dividual who have high fitness value is taken by the ture selection on input attributes. The parameters of
attribute. GA are shown in the following.

(4) The number of individuals : 40


The number of generations : 40
where c denotes a gene index, that is, the information
of an attribute. Ng is the maximum number of genera- The probability of mutation : 0.05
tions decided in advance. Let ak,c,min is the minimum
a k , c . Then c (attribute) in the next generation is set
The results of all the methods are shown in Table 1,
to 0. Table 2 and Table 3.
In GA, the individual of the next generation are de-
cided by the elite strategy until the Ng/2 generation. Table 1 Results obtained using the B P method.
After the N g / 2 generation, the individual is generated
7-
according to the Gaussian deisribution N ( U ~ , a2).
That is, the probability which a gene value is within
~,,
7Recognitionaccur;icyforfogI
( a k , c , - 0,a k , c , + 0)is 68,2%, ( a k , c , - 20,ak,c, + 20) Recognition accuracy for non-fog
+
is 95.5% and CL^,^,, - 3 a , a k , c n 30) is 99.7%. Here,
0 is the standard deviation here.
From these procedure, the priority of an attribute Table 2 Results obtained using the (LVQ1 and
can be obtained by using a real-coded GA in this pa- LVQ3) method.
per. Then, inputs in consideration of the attribute
priority rate can be chosen. Therefore, it becomes pos- I L V U l t LVQ3 I
I
sible to get a higher recognition rate than conventional Recognition accuracy or og
LVQ method. Furthermore, the number of inputs can
be reduced by selecting the attribute which isn’t nec-
essary for the recognition. This deduce a memory ca-
pacity.
Table 3 Results obtained using the proposed
4 Computer simulations method.
In order to verify the effectiveness of the proposed
scheme, computer simulations were done.
The learning data are made over into new data. In
this case, the number of the “fog” data and that of
“non-fog” are the same. Furthermore, the new term
(tempreture - dew-point) is added to the input at- From these results, the proposed scheme works better
tributes. than the other methods.

3680
5 Conclusions
In this paper, we propose a new method to predict a
fog occurrence by using a LVQ and a GA. The data
were provided from the Japan Meteorological Agency
and there is the one for 12 years. However, these data
contains an enormous amount of information. In this
paper, only inputs which are necessary for the predic-
tion are chosen by using a GA. Then, the time for the
fog occurrence forecast is shortened. Furthermore, in
order to show the effectiveness of the proposed scheme,
we performed a computer simulation for the fog occur-
rence forecasting.
The features of the proposed scheme are summa-
rized as follows:
0 The proposed method can make a small-sized sys-
tem. Therefore, the time for the fog occurrence
forecast is shortened.
0 The suitable LVQ network inputs can be obtained
by using a GA.
0 The importance of the attributes and the rela-
tionship among them can be obtained by using
Real-corded GA.
0 Recognition accuracy close to 100 percent can be
obtained.
0 Rule generation from a small-sized network would
be performed.

References
[l] T.Kohonen : Self-Organezed Maps, Springer-
Verlag Berlin Heidelberg (1995)
[2] T.Kohonen : In Advanced Neural Networks, ed.
by R.Ekkmiller (Elsevier, Amsterdam, Nether-
lands 1990)
[3] H.Freeman : IRE Trans. EC-20 (1961)
[4] D.E.Goldberg : Genetic Algorithm in Search
, Optimization & Machine Learning, Addison-
Wesley (1992)
[5] H.Kitano : Genetic Algorithm, Industrial li-
brary(1993) (in Japanese)
[6] I.Ono, H.Satoh and S.Kobayashi: A Real-
Coded Genetic Algorithm for Function Optimiza-
tion Using the Unimodal Normal Distribution
Crossover, Japanese Society for Artificial Intel-
ligence, Vol. 14, No.6 (1999)
[7] L.Davis : The Handbook of Genetic Algorithms,
Van Nostrand Reinhold, NewYork (1990)
[SI M.Fukumi and N.Akamatsu : A New Rule Ex-
traction Method from Neural Networks, Proc. of
IJCNN’99, #713 (1999)

3681