6.3 Experimental results
Because the goal is to compare the performance of
DE parallel implementations running on the GPU
the experimental results were registered for both
algorithms (sequential and parallel) and for each
tested function. In this section the observed
behaviors for DE implementation, after varying
iterations and individuals number, are commented.
Since it is well known that the convergence quality
of a given population-based algorithm, during an
optimization process, typically is very sensitive to
its specific parameters (either CR and F for DE)
and on the problem itself (i.e. objective function),
determining were DE algorithm has a better
convergence for each kind of problem is out of the
scope of this work. Regardless DE convergence
respect to a particular function, it is more
interesting in determinate the effect of code
parallelization itself and also the effect of varying
individuals and iterations number on both the cost
(i.e. consumed time) and the general shape of
convergence curve.
All tested functions have an optimum value at zero
except for F04. In order to have comparable plots,
the plot of F04 was adjusted by just adding the
optimum value (12569.4866 with 30 dimensions) to
show a convergence curve referred to zero value.