|
Adrian Horia Dediu; Lars Hildenbrand;
Claudio Moraga
A Remote Objects Implementation of Distributed Evolutionary Algorithms for
Performances Analysis Abstract.
We designed and implemented a new Distributed Evolutionary Algorithms
architecture based on remote objects. For experiments we used a relatively large
test function set starting with simple problems and continuing with more
complicated test problems with many local optima. First we did a very large
number of experiments for a simple test problem and then we randomly selected
samples consisting in 100 experiments. We observed that the average number of
evaluations until the solution was found for the sample tests differ no more
than 4% by the average number of evaluations for the whole test set. We tested
the behavior of Evolutionary Algorithms running in similar conditions in
centralized mode, distributed isolated evolutions with stop message between
sub-populations and distributed with individual exchange. The results of the
tests for maximizing the threePeaksExt function showed that Distributed
Evolutionary Algorithms perform better than centralized Evolutionary Algorithms,
mainly due to the fact that premature convergence observed in centralized
Evolutionary Algorithms is delayed due to individuals exchanges between
sub-populations. |