Information Technology Reference
In-Depth Information
Consider two chromosomes C 1 =( c 1 , ......, c n )and c 2 =( c 1 , ......, c n ). Ap-
plying the crossover operator on C 1 and C 2 generates two offspring H 1 =
( c 1 , .., c i ,c i +1 , ...., c n )and H 2 =( c 1 , .., c i ,c i +1 , ...., c n ).
The mutation operator is used to foster more exploration of the search space
and to avoid unrecoverable loss of genetic material that leads to premature con-
vergence to some local minima. In general, mutation is implemented by changing
the value of a specific position of an individual with a given probability, denom-
inated mutation probability. We developed new mutation operators tailored for
our genome representation that Shift the starting location and/or the ending
location towards the right or the left: the value in the starting/ending location
of a randomly selected gene is increased or decreased by a fuzzy fraction of the
upper and lower limits. This is implemented in such a way that the lower and
upper bounds of the domain of the field are never exceeded.
2.3 Fitness Function
As each chromosome in our approach comprises one rule, the fitness function
actually measures the collective behavior of the rule.
Definition 1. Accuracy of a certain rule is the ratio of correctly classified in-
stances by this rule to the whole number of instance to which this rule satisfied
f 1 = C
n
= Number of correctly classified instances by this rule
T otal number of instances satisfy this rule
2.4 Ending Criteria
As the GA searches the solution space, defined by the representation of the
application at hand, some of the rules are selected in the second step as promising
candidate solutions based on their evaluated fitness. They will undergo genetic
operations to generate a new population. This process is repeated until the end
criteria is met; either that the algorithm has reached a good enough solution or
that the maximum number of generations have been iterated. We use a hybrid
ending criteria that exits the execution whenever a good solution is met (human
expert defined) otherwise the execution will continue after maximum number of
generations.
2.5 Class Decomposition
Classification techniques usually assume homogeneity of class representation (the
number of cases represent each class is close); therefore the absence of a class in
the result classifer due to the trivial existence of that class in the training data is
possible. Our technique addresses this problem by building a separate classifier
for each class and then integrate them into one classifer. We used the training
data of all classes to calculate the accuracy of each classifer (for each class) to
Search WWH ::




Custom Search