Information Technology Reference
In-Depth Information
Table 9.4
Summary table of results for best genomes
Evaluation methods
BEST genome
Electrode selected
Frequency selected (Hz)
CCR (%)
CART
F4
1/8, 1/4, 2, 4 et 64
89.33
SVC
F2
1/32, 1/16, 2, 4, 8, 64 et 128
89.33
Figure
9.22
gives the occurrence of the electrodes in the best genome over the
100 runs. When some genomes have the same CCR at the end of the run, we select
the electrode chosen most often among the genomes with equal CCR. The algo-
rithm running with CART selects the electrodes around the number 10 (FZ in
Fig.
9.1
), 17 (FC1), or 30 (T4). With the SVC method, the electrodes around the 2
(FPZ), the 11 (F2), or the 48 (T6) are mostly chosen. Finally, on average, the
population of the evolutionary algorithm converges in less than 50 iterations for
both methods. Figure
9.23
gives the number of differences among parents for one
run of the algorithm. It shows that the number of differences among parents
decreases very rapidly and falls below the threshold of 3 differences in less than 40
iterations. Then, one of the three stop conditions is satis
ed and the algorithm stops.
Tables
9.3
and
9.4
summarize the CCR obtained by the GA, which are better
than those obtained (see Fig.
9.17
) with the criterion of the slopes calculated for
frequencies between 4 and 16 Hz (alpha waves). Moreover, Table
9.5
shows that
the GA allows for a dimension reduction. SVC classi
er cannot be used with more
than one variable. Then, Table
9.5
only shows a comparison between the results
obtained in Sect.
9.5.3
and those obtained with the GA for the CART classi
er.
It also appears that it is more appropriate to use a regression on frequencies of
1/8, 1/4, 2, 4, and 64 Hz for the signal of electrode
F
4 and the CART classifier.
Then, this work allows to accurately predict the state of alertness of a new indi-
vidual. In fact, this electrode and this range of frequencies will be used to calculate
the slope criterion for this individual. The CART decision tree, built on the sample
formed by the 26 signals (13 study participants in both states of alertness) will be
used as a classi
er to predict his state of alertness.
Fig. 9.22
Occurrence of the
electrodes in the best genomes
for each electrodes during the
100 runs of the GA with 300
parents, 150 children, and
CART (
dash-dotted curve
)or
SVC (
solid curve
)
50
CART
SVC
45
40
35
30
25
20
15
10
5
0
0
10
20
30
40
50
60
Electrode number
