Information Technology Reference
In-Depth Information
duplicates of their predecessors the evolution of the population depends on muta-
tion of the clones.
4.
Elitism:
For population-based GA, during each generation, the whole population is
replaced with the offspring after mating; so 'elitism' has to be introduced to pre-
serve good solutions found so far, otherwise they would be lost during generations;
for AIS, the mutated clones and their predecessors are mixed together to compete
for survival, so the 'elitism' is inherently embedded in AIS
5.
Population control:
For the population-based GA, since one has to specify the
size of the mating pool in the first place the population size is thus fixed during
each generation; if one only selects good solutions into the mating pool and makes
the pool size flexible to the number of selected solutions GA could end up reaching
premature convergence due to its evolutionary strategy; a reasonable pool size is
necessary so that in the early stages sub-optimal solutions can also be included in
the pool to increase population diversity; for AIS, a mating pool does not exist
hence the population can be flexible and finally controlled by the mutual influences
of Abs.
5.2 Extra Strength of AIS
If one recognizes all these differences AIS should offer extra strengths, which have
been implemented in PAIA and are summarized as follows:
1.
Adaptive population.
Network suppression was first proposed in [3] to perform
data analysis. In PAIA, it is used to regulate the population. The main point is that:
it allows any selected Ab to get into to the population as long as it is far enough
from any other Ab. This flexible rather than fixed population plus adaptive clone
size make the population adaptive in the problem.
2.
Initial population size is not crucial to the success of PAIA.
Due to the nature of
the adaptive population, whatever initial size is used the population can be adap-
tively adjusted to a reasonable size according to the need of the problem. Although
the results are not shown in this paper, in other experiments it was found that one
can use any number as the initial population size (even 1) and the results in terms
of performance metrics are equivalently good as the ones presented in this paper.
The only difference is if an optimal initial size is chosen the evaluations can be
largely reduced.
3.
Fast convergence.
In PAIA, even a small initial size (e.g. 7) can give a very fast
convergence because one only selects good Abs and let them reproduce with an
adaptive clone size. In the early iterations this cannot only provide sufficient Abs
to support the search but also accelerates the convergence.
4.
Only necessary evaluations are exercised.
Since only a necessary population size
and clones are maintained and produced in each iteration step, only necessary
evaluations are carried out. One can see from Table 5 that PAIA used 46899
evaluations to converge for ZDT4. If one uses the same setting for NSGA II (a
population of 100 and 500 generations) 50000 evaluations would be needed.
5.
Parameter less.
The only parameter crucial to the success of PAIA is the way to
calculate the mutation rate. However, an adequate combination of parameters can
efficiently tackle this problem (using a fewer evaluations).
