Biomedical Engineering Reference
In-Depth Information
Table 1. IRIS: Results of ANN training only with GA
Generation
MSE
% Test
Seconds
11
1
0,42
50
0:00:01
21
0,39
58
0:00:01
2
3
0,34
48
0:00:06
117
4
0,37
60
0:00:00
7
1455
5
0,25
54
0:01:17
0,33
56
0:00:01
4
6
13
7
0,39
54
0:00:01
19
8
0,38
58
0:00:01
9
0,4
62
0:00:01
11
10
0,44
60
0:00:00
8
Table 2. IRIS: ANGN results of the ten populations
G eneration
MSE
% Test
S econds
44
1
0,09
82
0:00:12
2
0,04
76
0:03:28
780
1045
3
0,06
80
0:04:31
1001
4
0,02
84
0:06:15
5
0,07
82
0:01:17
297
6
0,04
80
0:02:33
404
1516
0,09
82
0:06:32
7
8
0,08
80
0:01:28
316
51
9
0,07
64
0:00:13
10
1485
0,09
72
0:06:25
test phase. Although training time was shorter
in the case of ANN training only with GAs, it
is important to emphasize that the simulations
continued until 4000 generations in order to unify
the results; also the percent accuracy of the test
with ANNs did not reach more than the indicated
value during the 4000 generations.
Considering the results obtained so far with all
the simulations, it was checked the effectiveness of
the chosen options of artificial astrocyte behaviour
for the non-supervised stage. The reason why
these satisfactory results were achieved with the
ANGNs is attributed to what was expected when
the hybrid method was designed. The individuals,
once evaluated, are presented to the GAs arranged
by their MSE. The changes made to the weights
during the unsupervised learning phase cause this
resulting order to be different from a potential
order obtained in case this phase did not exist.
This is so because their MSE varies according to
this new methodology. The process of searching
for the global minimum carried out by the GAs
adds to the local adjustment process implemented
thanks to this non-supervised stage.
Search WWH ::
Custom Search