Geography Reference
In-Depth Information
are usually rather long and complex and adaptive genetic operators are needed
(Janikow 1993 ). On the other hand, the Michigan approach generates simple and
short chromosomes and has the advantages that the classification rules can be easily
evaluated and the normal genetic operators can be applied. Nevertheless, since each
rule is generated separately, it is necessary to run several times the algorithm to
obtain all the classification rules. Furthermore, quality of the classification rules as a
whole is hard to evaluate (Greene and Smith 1993 ). Based on the above discussions,
the choice of which approach to use mainly depends on which types of classification
rules one wants to obtain. Further, if the quality of the rules as a whole rather
than that of single rules needs to be evaluated, the Pittsburgh approach seems more
suitable, and vice versa (Freitas 2002 ).
13.3.2.2
Definition of Fitness Function
An important step of the GA is to evaluate the classification rules or chromosomes
and based on which determine the chance of survival of the chromosomes into
the next generation. Freitas ( 2002 ) suggests that the fitness of the chromosomes
should be evaluated by their predictive accuracy, which can be measured by the
product of the so-called confidence factor and completeness factor. The confidence
factor is the percentage that the predictions of the chromosomes are correct, whereas
the completeness factor refers to the percentage that the actual cases are correctly
predicted by the chromosomes. Noda et al. ( 1999 ) proposed a fitness function
measuring not only the predictive accuracy but also the degree of interestingness
of the classification rules. Given that the user would tend to be more surprised and
interested when he/she saw some seemingly irrelevant attributes were attributed into
relevant ones, the computation of interestingness value is based on the following: the
smaller the probability of the goal attribute value, the more interesting it is. Fitness
function is essentially a weighted sum of two terms measuring predictive accuracy
and degree of interestingness and the weight assigned to each term is set by user.
13.3.2.3
Genetic Operators
Chromosomes with high fitness values are selected from the population to be parents
for crossover and mutation operations. According to Darwin's evolution theory,
the best ones should survive and generate new offspring. Several methods have
been proposed to select the best chromosomes for generating offspring including
roulette wheel selection, rank selection, tournament selection (Razali and Geraghty
2011 ). The Roulette Wheel Selection, which is also known as fitness proportionate
selection, allocates probabilities to be selected to the chromosomes in proportion
to their fitness values. An advantage of this method is that all individuals in the
population are given a chance to be selected so that diversity in the population
Search WWH ::




Custom Search