Biomedical Engineering Reference
In-Depth Information
Figure 7. Typical IR spectrums
S amples
0. 15
0. 1
0. 05
0
20
40
60
80
100
120
140
160
wavenumber
in detail it was observed that its spectrum was
slightly different from the usual ones. This would
suggest that the juice contained an unusually high
amount of added sugar(s).
by the different approaches is to find out a small
set of IR variables that, when combined with an
ANN model, might be capable of classifying apple
juice-based beverages properly (according to the
amount of pure apple juice they contain).
Once this subset of IR variables is proposed,
their associate absorbance values are used as input
patterns to the ANN. So, ANN will consider as
much input processing elements (PEs) as variables.
The output layer has one PE per category (6 for
the lower range and 5 for the higher concentra-
tions). After several previous trials considering
several hidden layers (from 1 to 4), each with
different PEs (from 1 to 50), a compromise was
established between the final fitness level reached
by the ANN and the time required for its training.
This compromise was essential because although
better results were obtained with more hidden
layers the time required for training was much
higher as well. It was decided not to extensively
train the network but to get a good approximation
to its real performance and to elucidate whether
the input variables are really suitable for the ac-
curate classification of the samples.
variable Selection Approaches
Once the data are available, a variable selection
process can be performed in order to optimise
the ANN based model.
First, two simple approaches will be briefly
described: pruned and fixed search. Both ap-
proaches are based on a traditional GA and use
ANN for fitness calculation. As the results show,
both techniques offer good solutions but present
a common problem: an execution of each method
provides only a valid solution (discarding the
optimal ones). Later, this will be addressed with
the two approaches described in the previous
sections.
All the approaches use a GA to guide the search
by evaluating the prediction capabilities of differ-
ent ANN models developed employing different
sets of IR variables. The problem to be addressed
Search WWH ::
Custom Search