Information Technology Reference
In-Depth Information
After applying to these values a collection of mathematical functions at different
stages, an output value is given. This output value has a useful meaning to assign
a class to an input sample (e.g. positive or negative for cancer), or express the
probability for this sample of being a member of a class. A classification of ar-
tificial neural networks based on the kind of learning approach followed in their
design, divides them into two types: supervised and unsupervised networks [15].
The supervised networks need a supervision of the automatic training process,
which specifies the corresponding output to each training input. In the case of
unsupervised networks, it is expected that the network classifies the inputs into
different groups according to the outputs, but in this case it is not explicitly spec-
ified to them the relationship between input and output. In this document, we
will work with two different supervised ANNs: multi-layer perceptrons (MLPs)
and radial basis functions neural networks (for short RBFs).
In the case of the MLP implementation, it is mainly used the backpropagation
training algorithm to carry out this process [15]. MLPs have an input and an
output neuron layer, and a non-fixed number of intermediate, or hidden, layers
with an arbitrary number o neurons in each one.
In the case of RBF's, its structure is quite similar to multi-layer perceptron,
with an input and output layer respectively. However there is a difference in
the number of intermediate, or hidden, layers. RBFs have only one hidden layer
while MLPs could have more than one. In addition to this, another important
difference between MLPs and RBFs is the transference function that each neu-
ron applies to the data that pass through them. In the case of MLPs it is a
linear function, in most cases a sigmoid, and in the RBFs case in the intermedi-
ate layer it is radial function, mostly any kind of gaussian function, and in the
output layer a linear function too.
3 Experimentation
Next, we explain the steps followed in the design, implementation and testing
processes of each of the automatic classifiers. It is also presented the clinical
tests database used and the preprocessing data tasks that have been carried
out. Finally, we explain the results obtained from the test phase for each ANN
and the GA implementation, comparing them in terms of accuracy rates.
3.1 Urological Tests Database
The database used to carry out the tests consists of samples of patients with
urological problems. Our objective has been to diagnose the two main diseases
that these patients suffer from. All along the experimentation we were advised
by medical specialists, who helped us to have a clear idea about the meaning of
the data included in the clinical tests.
Our database consists of 160 samples of patients, suffering 74 of them from
incontinence and 86 from urinary tract obstruction. Each sample consists of
Search WWH ::
Custom Search