Information Technology Reference
In-Depth Information
Table 9.2
Logical operator used for the frequencies during the crossover
Parent 1
Parent 2
Child
0
0
0
1
Bern
ð
1
=
2
Þ
0
Bern
ð
1
=
2
Þ
1
1 1
For a binary component, when the two parents do not share the same value, a Bernoulli
distribution is used to build the component of the children
0
1
When the two parents do not share the same value, then a Bernoulli distribution is
used to build the component of the children.
Mutation
Once the child is established, a mutation is applied. Each component of the genome
of the child mutates with probability 1/8. Thus, each child is, on average, affected
by two mutations. When a mutation is applied to the electrode number, a random
number (drawn between 1 and 58) replaces the current value. For the binary part, a
mutation is a bit-
ip operation (a 0 becomes a 1 and vice versa).
9.6.2.2 Evaluation Functions
The GA searches for the best combination of electrode/frequency range which
achieves the highest prediction accuracy. Thus, it seems natural to rely on the CCR.
Then, the
fitness function corresponds to the CCR obtained for each genome. These
are then ranked in descending order of CCR. To compare each genome, the same
samples are used to calculate the CCR using a
fivefold cross-validation. The
evaluation step is done for each child at each iteration. Thus, it is necessary to use a
fast classi
cation method as evaluation function. In this work, two methods have
been tested (see algorithms 1 and 2). The
cation
(SVC) (Guyon and Elisseeff,
2003
), a method to predict from a single variable. The
average for each modality (normal or relaxed) is calculated on the individuals in the
training set for the variable (feature). Individuals of the test sample are then
assigned to the class corresponding to the nearest average. The prediction is
compared to ground truth which gives a CCR. The second method is the binary
decision tree (CART) (Breiman et al.
1984
). Here, the algorithm is used with a
single variable which guarantees fast calculation. Then, the
first is the single variable classi
fitness function for each
genome
x
is written as:
f
ð
x
Þ¼
#
well classified participants of the test set
#
participants in the test set
ð
9
:
4
Þ
