Recent Methodology in Connectionist Systems - Advancing Artificial Intelligence Through Biological Process Applications - page 100

weights do not consider the error of the output, but
take place at any time according to the activation
frequency of each neuron, simulating reinforce-
ments and inhibitions in the brain circuits.
The second and last learning phase consists
of applying GAs to the individuals according to
their MSE, which was stored in the first phase.
The GAs phase carries out the corresponding
crossovers and mutations and selects the new
individuals with which the first and second phases
will be repeated until the least possible error, or
preferably no error, is obtained. The second phase
is considered a supervised training because the
GAs takes into account the error made by the net-
work to select the individuals that will be mutated
and crossed-over, i.e. it makes the changes in the
weights according to that error.
During the testing phase (once training has
finished) it is necessary to test the ANGN gen-
eralization capacity. If this capability is correct,
ANGN will be ready for its subsequent use. We
want to emphasize that at this phase, and in the
course of all the subsequent runs, the brain be-
haviour introduced in the non-supervised learning
phase will always be applied, since it is a part
of the ANGN in all its stages and it participates
directly in the information processing. The input
patterns will be presented to the net during the
iterations that were determined in the training
phase, which allow the ANGN to carry out their
activity.

can be noticed that, in most of them, the artificial
astrocytes must induce network modifications
based on neuronal activity. With our work we
have tested the first proposed options, related to
the connections between neurons. Such options
were implemented, as it has been mentioned, at the
non-supervised stage. The efficacy of the non-su-
pervised stage was shown by means of comparing
the ANGNs with the corresponding ANNs trained
only by means of GAs. We made said comparison
at solving two classification problems of various
complexities. The tests started with a very simple
problem: MUltipleXor (Porto et al., 2005). Later,
in a more complicated domain, initial tests were
performed with the IRIS flower problem (Porto et
al., 2007). Such tests have been recently completed
with multiple simulations whose summary results
are following shown.

IRIS Flower Problem

This problem has served to test the ANGNs func-
tioning when dealing with multiple classification
tasks. IRIS flower problem uses continuous input
values, different activation functions in artificial
neurons of different layers, and twice as many
training patterns. This example has been carefully
analysed in the field of ANNs since A. Fisher first
documented it in 1936 (Fisher, 2006). It consists
on identifying a plant's species: Iris setosa, Iris
versicolor and Iris virginica. This case has 150
examples with four continuous inputs which
stand for 4 features about the flower's shape.
The four input values represent measurements
in millimeters of the: petal width, petal length,
sepal width, sepal length. We have selected one
hundred examples (33,3% of each class) for the
training set and fifty examples for the test set, with
the purpose of achieving a great generalization
capability.
The learning patterns have been found to
have 4 inputs and 3 outputs. The three outputs
are boolean ones, representing each Iris species.
By doing it in this manner (three boolean outputs

LATEST RESULTS

The analysis of the cerebral activities has opened
various ways to convert ANNs into ANGNs and
provide them with a potential that improves their
contribution to the information processing. We
established possibilities of conversion that were
classified by us according to what happens with
connections between neurons, the activation value
of the neurons, and combinations of both (Porto
& Pazos, 2006). By analysing these options it