Neural Networks versus Genetic Algorithms as Medical Classifiers - Foundations on Natural and Artificial Computation - page 397

Fig. 1. Configuration parameters of the MLP final implementation
A similar approach was followed for the RBF implementation. In this case the
tested parameters were the radial basis function to be used as transfer function
in the intermediate layer, the ridge of this function and the minimun standard
deviation allowed (Fig. 2).
Fig. 2. Configuration parameters of the RBF final implementation
For the design and implementation of the classifier based on genetic algo-
rithms, we tested the combination of different values for each parameter of the
algorithm: initial population size, maximum number of generations to be created,
target function for selecting the chromosomes, type of intersection between elite
chromosomes etc.
The first step was to obtain a set of representative values of each class that could
be assigned to the input samples. To do this, we generated randomly an initial
population where each chromosome has an associated feature vector. Next step
is selecting the elite chromosomes to do the crossing that generates the following
generation. The elite chromosomes are selected using as deciding factor, the dis-
tance between their features vector and the training samples set for each class. The
objective function which value we want to optimize, in this case minimize, it is a
distance function, so we tested three classic distance functions such as Manhattan
distance, Normal or Minkowski and Euclidean. It was the latter which gave the
best results and therefore was the chosen one for the final design.
After obtaining the representative patterns of each class by means of the GA, the
next step was to calculate the value of the Euclidean distance from each sample of
the testing set to these class identifying patterns. Finally, it was assigned the class,
i.e. the diagnostic result, of the set that was closest to each sample of the testing
set, that is, the one that makes the distance function value minimum.
The combination of the GA configuration parameters that provides a bet-
ter result in the diagnostic process, finally imlpemented, is showed below in
(Fig. 3):
Fig. 3. Configuration parameters of the GA final implementation