Genetic Algorithms and Multimodal Search - Advancing Artificial Intelligence Through Biological Process Applications - page 235

the individuals that exploit unique areas of the
domain. The dynamic fitness sharing (Miller &
Shaw, 1995) was proposed in order to correct the
dispersion of the final distribution of the individu-
als into niches with two components: the distance
function, which measures the overlapping of
individuals, and the comparison function, which
results “1” if the individuals are identical, but gets
closer to “0” as much different they are.
The clearing method (Petrowski, 1996) is quite
different from the previous ones, as the resources
are not shared, but assigned to the best individuals,
who will be then kept at every niche.
The main disadvantage of the techniques previ-
ously described lies in the fact that they add new
parameters that should be configured according
the process of execution of GA. This process may
be disturbed by the interactions among those
parameters (Ballester & Carter, 2003).

proposals that use EC techniques for this type
of problems. Both proposals try to find the final
solution but keeping partial solutions within the
final population.
The main ideas of the two proposals, together
with the problems used for the tests are explained
at the following sections.

Hybrid Two-Population Genetic
Algorithm

To force a homogeneous search throughout the
search space, the approach proposed here is based
on the addition of a new population (genetic pool)
to a traditional GA (secondary population). The
genetic pool will divide the search space into
sub-regions. Each one of the individuals of the
genetic pool has its own fenced range for gene
variation, so every one of these individuals would
represent a specific sub-region within the global
search space. On the other hand, the group of
individual ranges in which any gene may have its
value, is extended over the possible values that a
gene may have. Therefore, this genetic pool would
sample the whole of the search space.

Own Proposals

Once detected the existing problems they should
be solved, or at least, minimised. With this goal,
the Artificial Neural Network and Adaptive
System (RNASA) group have developed two

Figure 2. Structure of populations of hybrid two-population genetic algorithm

0

…

Ind 1

G 11 G 12 G 13 G iN

d/n

d/n Ind 2

…

G 21 G 22 G 23 G iN

N
Number of Variables Selected
P
Genetic Pool Individuals
S
Secondary Population Individuals
d
number of original variables
n
number of subregions to divide
the search space

2·d/n

(n-1)d/n
d

…

G P1 G PN Ind P

G P2 G P3

d

Ind 1

…
…

0

Ind 2

d

0
d

…

Ind S