Biomedical Engineering Reference
In-Depth Information
Figure 6. Codification of the activation functions. (a) Hyperbolic (b) sigmoid
Activation F.
Hyperbolic
Activation F.
Hyperbolic
Activation F.
Hyperbolic
Activation F.
Activation F.
Activation F.
Parameter
Parameter
Parameter
a)
b)
Hyperbolic
1
1
Sigmoid
Threshold
Lineal
Hyperbolic
Sigmoid
Threshold
Sigmoid
Threshold
Lineal
Hyperbolic
Sigmoid
Threshold
Sigmoid
Threshold
Lineal
Hyperbolic
Sigmoid
Threshold
Sigmoid
Threshold
Lineal
Hyperbolic
Sigmoid
Threshold
1.0
0.85
1.0
0.85
1.0
0.85
1.0
0.85
1.0
1.0
1.0
1.0
0
0 .5
0.85
0.85
0.85
0.85
- 1
1.5
1.5
1.5
1.5
0
Figure 7. Genetic Algorithm used
Initial
Population
Initial
Population
Initial
Population
Initial
Population
Initial
Population
Initial
Initial
Initial
Initial
Initial
Population
Population
Population
Population
Population
Selection
Selection
Selection
Selection
Selection
Crossover
Crossover
Crossover
Crossover
Crossover
Mutation
Mutation
Mutation
Mutation
Mutation
the
Population
the
Population
the
Population
the
Population
the
Population
Sort
the
Sort
the
Sort
the
Sort
the
Sort
the
Sort
Sort
Sort
Sort
Sort
END
END
END
END
END
END
END
END
END
END
MSE <
MSE <
MSE <
MSE <
MSE <
Population
Population
Population
Population
Population
individuals and two crossover operators were
applied. In each of these two operations, the
crossover operator behaves in the usual way: a
divide point is chosen at random and the new
individuals are generated. The result of these
two crossover operations are two new individuals
with a weight part and a line gradient part from
each parent. Then the new individuals are evalu-
ated in order to apply the individual substitution
strategy (Figure 9).
Apart from the usual crossover operator with
one crossover point that has just been described,
tests have been carried out with a two-point
crossover operator and with the uniform operator
which selects genes from one of the parents in each
position according to a uniform distribution.
Once the new individuals have been obtained,
it only remains to choose which individuals in the
population are going to leave their place to the
new ones. Apart from the usual and more Dar-
winist technique of individual substitution, which
eliminates the worse adapted ones, two additional
techniques have been tested in order to avoid ho-
mogenization. The first one is about substituting
parents, i.e. if the offspring adapts better than the
parents, these are substituted, in the opposite case,
the offspring is eliminated. The second technique
is based on a substitution according to similarity
of error level. Those individuals with an equal or
similar error level to the new offspring are found
among the population, and one of them is chosen
at random to be substituted. In the Figure 10 we
can see all the parameters for the implemented
Genetic Algorithm.
Also, it's codified if the ANN is recurrent or
not and the generic parameters of the PE, as the
Search WWH ::
Custom Search