|
|
i
i
il
I
|
|
|
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
E(L;)= E(L j)
MI Wpositior position Li, j) IE Li j without conflicEO
0 otherwise
; N
Ji(1)s Y EG) (7)
i=]
Where W esiti on l.
Now consider the example in figure 2, the figure gives out its
chromosome. Consider the targets of least conflict and most
optimal position, adopt the fitness function (7), and calculate its
fitness value:
20
fi(L) 2 X E(Lj) 2 99x10 98x44 97x3--96x3 —1961.0
i-l
4.4 design genetic operators
4.4.1 Selection operators
The labeling algorithm adopts roulette-wheel selection as
selection method. When the population size is very large, can use
elite strategy, namely retain the optimal individuals of the
previous generation into the next generation's population. The
implementation algorithm of roulette-wheel selection is as follow:
(1) Calculate the fitness value of individual
fit i =12,...,n);
(2) Calculate the accumulative fitness value of individual
Accfit(V, Yi = K2 m) and relative accumulative fitness
value Re/Accfit(V, Xi = 1290).
(3) Generate a random r in [0, 1],
here suppose Re lAccfit(V,) zi,
Re/Accfi(V, ,) « r x Reldccfi(V,) (i=12,.n),
then select the individual i.
4.4.2 Crossover operator
(l)single-point-crossover:
father string1 (0[1]2]3]0]1]0)2
Iu NS poin exchange _ =
father string2 TORENT 1J0|1]1]2|2]1]0
offspring string?
offspring stringl
(Zymultipoint- crossover:
father stringl O]1/2]5]0] 11012 : y Ont]1H1[2]1]0/2
; er pointl Y T zm: point2 offspring stringl
father string? [1|0111]212]1|0 110121310121 110
exchange offspring string?
Figure 3 the example of point-crossover
The integer vector generally adopts two kinds of crossover
operators: point-crossover and even-crossover. The experiments
have proved there is no too much difference between the
influences of two kinds of crossover operators on the
performance of labeling algorithm. Therefore the labeling
algorithm adopts the point-crossover strategy.
The point-crossover operator is divided into single-point-
crossover and multipoint-crossover. The single-point-crossover
randomly selects a cross point on two father strings and then
exchanges the corresponding sub-strings of the two strings. The
multipoint-crossover randomly generates several cross point
every time, and then exchanges the corresponding sub-strings of
father strings respectively [Pan Zhengjun, 1998]. Figure 3 gives a
chromosome point-crossover example of two labeling place-
ments with the string length 8 and the gene code set 10, 1, 2, 3}.
4.4.3 Conflict gene mutate (namely mutate on conflict gene)
For integer vector encoding, the common mutation includes
point-mutation and even-mutation. The former selects single
point, the latter selects point according to some template, and
then randomly relocate the selected point.
As for map labeling, we put forward a new mutation operator
which is called conflict gene mutation to replace the routine
point-mutation and even-mutation. The basic theory of conflict
gene mutation is that: select the gene of conflict label, randomly
generate a labeling code to replace the original gene.
Experiments have proved the conflict gene mutation is very
effective. Figure 4 shows the comparison result of the conflict
gene mutation and even-mutation with the same genetic
parameters (the number of iterations is 300, population size is 50,
crossover probability is 0.75, mutation probability is 0.2). From
figure 4 we can find that the conflict gene mutation is obviously
superior to the even-mutation. The possible reasons are as
follows:
(1) The point-mutation and even-mutation are blind, and the
conflict gene mutation utilizes the heuristic information of label
conflict to improve the bad sub-solution, so the probability of
obtaining good sub-solution is bigger.
(2) In actual maps, there is little conflict in the area with sparse
labeling point, the mutation probability should be smaller,
however in the area with dense labeling point there is more
conflict, and the mutation probability should be relatively larger.
The conflict gene mutation meets this demand.
1 — & E E e
qd ptt "s
o 0355! +
5 |
Ca — 09! +
a |
0.85:
E à | e +
= + Mutate on . |
= 0.75 + Conflict Gene
3, E : 5
s ds | Even Mutate
"rj & i +
o 0.65 |
E nl
u i *
5 0.55
5S 09
ü 200 400 600 200 1000 1200 1400 1600
Labeling Poing Numbers
Figure 4 compare conflict gene mutation with even mutate
4.5 determine the important parameters of genetic algorithm
According to references and experiment conclusion, there are the
following principles in determining the genetic parameters of
automated labeling:
620
Inter
(1) 1
algor
doma
(2) X
Cross
this p
(3) M
the d
03.1
(4) T:
over
termi
guara
to jud
opera
stably
impro
V.
CON
The e
two p
by the
actual
topogi
1:250(
layers
The th
and 27
From |
point |
gradu:
genetk
has m
popula
point
MILYSAE.
‚#8
ET
su
“14e
vaux
Figure
The ex
algoritl
compar
on the ¢
the pro
parame
populat
increase