Toward Sensor to Reproduce Human Sense - The Electronic Nose Artificial Olfaction Technology - page 136

regression (MLR), canonical correlation regression (CCR), multilayer perception
(MLP), radial basis function (RBF), probabilistic neural network (PNN), K nearest
neighbors (K-NN), support vector machines (SVM), adaptive resonance theory
(ART), genetic algorithm (GA), hierarchical clustering (HC).
The Pattern reorganization methods can be divided into unsupervised tech-
niques and supervised techniques. Unsupervised learning methods are generally
used in exploratory data analysis because they attempt to identify a gas mixture
without prior information on the nature of the samples. These techniques, which
include PCA, CA, ICA, SOM and MDS, are useful when no example of different
sample groups is available, or when hidden relationships between samples or
variables are suspected [ 19 , 20 ].
Supervised learning techniques classify measurements by developing a math-
ematical model to relate training data, i.e., samples with known properties, to a set
of given descriptors. Test samples are then evaluated against a knowledge base and
predicted class membership is deduced. These methods enable the system to de-
emphasize parameters other than volatile, for example temperature and humidity,
and to train a system to look only at particular combinations of sensors [ 19 , 20 ].
When high concentrations of volatile are involved, a non-linear PARC tech-
nique, such as ANN or RBF, would be more appropriate. Non-linear models
usually need more parameters and therefore more input data than linear models,
since some parameters are used to describe the shape of the non-linearity. The
main advantage of such a method is flexibility, i.e., the ability to adjust to more
complex data variations. However, caution is necessary when choosing the desired
model flexibility by selecting the number of parameters. If too many parameters
are taken into account, the calculated model will be over-flexible, fitting to all
relevant data variations as well as to every unwanted sensor noise.

• Dimensionality reduction:
A dimensionality reduction stage is required in most cases, either feature
extraction or feature selection.
Unsupervised: In unsupervised learning, the training of the network is entirely
data-driven and no target results for the input data vectors are provided.

- Multidimensional Scaling is specifically designed to graphically represent
relationships between objects in multidimensional space. The objects are
represented on a plot with the new variables as axes and the relationship
between

the objects on the plot should represent their underlying

dissimilarity.
- Principal components analysis is a signal representation technique that gen-
erates projections along the directions of maximum variance. PCA is a
method that reduces data dimensionality by performing a covariance analysis
between factors. As such, it is suitable for data sets in multiple dimensions,
such as a large experiment in gene expression. PCA transforms data so that
they are redefined in terms of their principal components (PCs). PCA is
powerful technique for data analysis.