Recent Methodology in Connectionist Systems - Advancing Artificial Intelligence Through Biological Process Applications - page 101

instead of a multiple one) additional information
can be provided about whether the system's outputs
are reliable or not. That is, due to the outputs'
intrinsic features, only one of them must possess
the true value, standing for the type of flower it
has classified, while the rest have a false value.
Therefore, if two or more outputs have true values,
or if all of them are false, we may conclude that
the value classified by the system is an error and
the system can not classify that case. The values
corresponding to the four input variables have
been normalized in the interval (0-1) so that they
are dealt with by the CSs.
We have started from optimal architecture and
parameters which were obtained by our research
group in previous works (Rabuñal et al., 2004):
5 hidden neurons, tangent hyperbolic activation
functions and threshold function (0.5) in the output
neurons. By using ANN with these features and
trained exclusively by means of GAs, J. Rabuñal
et al. (2004) reaching an adjustment better than
the previous best example of work for solving
IRIS flower with ANN, in which A. Martinez &
J. Goddard (2001) used BP for the training and a
hidden neuron more than J. Rabuñal et al. (2004).
These good results demonstrated the GA efficacy
for simplifying and solving this problem. We have
compared our ANGNs with these ANNs trained

exclusively by means of GAs taking in account a
maximum value for the weights of “1”.
The ANN architecture may be observed in
Figure 2.
The tests were carried out by keeping the
same ten populations of individuals and the same
random seed which originates the selection of the
individuals for crossover and mutation.
In order to draw an adequate comparison we
established the same GAs parameters for all the
tests: Population size of 100 individuals; Monte-
carlo technique in order to select the individuals;
Darwinian substitution method; Crossover rate
90% and mutation rate 10%; and a single cut
point for crossover.
We want point out that these GAs options are
not the ones which provide good results. Our
purpose was that said parameters coincide in the
CSs to be compared.
In this problem, we establish four thousand
generations for training the ANGNs in all the
simulations.
Table 1 shows the best results obtained for
the ten populations using the ANN training only
with GAs.
Table 2 shows the best results achieved with
the ANGNs for the same ten populations.
If we compare the two tables, ANGNs reached
minor MSE and major percent accuracy in the

Figure 2. ANN architecture for IRIS flower problem

I 1

Setosa

I 2

Versicolor

I 3

V irginica

I 4