Information Technology Reference
In-Depth Information
modules; where each module is responsible for developing a classifier for a specific
class. In other words, each class receives enough attention whatever the size of
the input data that represents that class. Accordingly, rare occurring classes
will not be ignored and still get their own attention. These modules are trained
in parallel and independently and the results obtained are integrated to a final
solution. Several benefits are gained using this class decomposition technique:
all classes are fairly represented; parallelism can be obtained for faster solution;
and scalability is possible as the problem is broken into several manageable sub-
problems. After this detailed introduction and related work review given in this
section, details about the methodology of class decomposition using a GA is
given in section 2. In section 3 four experiments are conducted to evaluate the
proposed technique. Finally, a brief conclusion is given in section 4.
2 Methodology
A GA is a search and optimization methodology from the field of evolutionary
computation that was invented by Holland [21]. A GA is based on the Darwin's
natural selection principle of the survival of the fittest, and is widely used for
hard problems in engineering and computer science. A GA is a population-based
method where each individual of the population represents a candidate solution
for the target problem. This population of solutions is evolved throughout sev-
eral generations, starting from a randomly generated one, in general. During
each generation of the evolutionary process, each individual of the population
is evaluated by a fitness function, which measures how good the solution rep-
resented by the individual is for the target problem. From a given generation
to another, some parent individuals (usually those having the highest fitness)
produce ''offspring”, i.e., new individuals that inherit some features from their
parents, whereas others (with low fitness) are discarded, following Darwin's prin-
ciple of natural selection. The selection of the parents is based on a probabilistic
process, biased by their fitness value. We use rank-selection mechanism which
means that a ranking process is performed on the fitness values then individuals
with the higher fitness value are selected for the next generation. The generation
of new offspring from the selected parents of the current generation is accom-
plished by means of genetic operators. This process is iteratively repeated until
a stop criteria is reached.
2.1
Individuals Representation
An individual in our method is a classification rule where each gene represents
the minimum and maximum values of intervals of each attribute that belongs to
such rule. In rule-based classification, there are various representation methods
in terms of rule properties (fuzzy or non-fuzzy) and attribute properties (nominal
or continuous). In our approach, we use fuzzy IF-THEN rules with continuous
attributes. A rule set (classifier) consisting of a - user determined - number of
Search WWH ::
Custom Search