Вы находитесь на странице: 1из 2

Text for RODDS paper

FIGURE GOES HERE


(Caption) Figure ? In each pane the symbol box around i means this is a decision
vector value evaluated in iteration i. I diamond around this symbol indicates in its
iteration it is the best solution, and from it the perturbations for the next generation
will be generated. The circles have radius R i and any perturbations falling inside
the circles are rejected.

Section An Overview of RODDS


RODDS is a new global optimization algorithm that we designed to effectively
search a space of sizeable dimension while using a significant number of processors
in parallel. Figure ? gives a visual description of the new RODDS algorithm. Figure ?
a represents search in a 2D space by serial DDS. In each iteration, only one new
point is generated and it becomes the new best solution only if it is better than the
previous solution. In Figure b there is a representation of the straightforward
parallelization of DDS called DDS-PC. In Figure ? b with two processors, two points
are generated in each iteration, and the points in iteration 2 are perturbations
around the best solution from iteration 1. Then from Figure ? b we see that the best
solution for iteration 2 with DDS-PC (from which the next perturbation will be made)
is quite close to the previous best solution. By contrast in Figure ? c, with the new
parallel method RODDS, there are circles of radius R around each of the previously
evaluated points. In RODDS any new perturbation that falls inside the circles is
rejected and an new random perturbation replaces it. So the X through the one of
the perturbations (denoted box around 2) indicates that perturbation is rejected
and is replaced by another point that is outside all the circles. In Figure ?d, we see
one additional iteration of RODDS that shows that the radii of the circles are
decreasing, meaning that new best points are allowed to get closer to previous
evaluations as iterations progress.
The purpose of the circles in RODDS is to prevent all the searching being in one
region when the computation is done with a sizeable number of processors in
parallel. In the serial DDS this is not a problem because only one new perturbation
is generated from each previous best point. However, if there are 32 processors, in
DDS-PC, there will be 32 perturbation points generated around each best solution,
which leads to a large number of function evaluations being on average close to the
previous best solution and does not encourage searching the whole space. This
problem might not appear to be too severe in the two processor, two-dimensional
case above, but with a higher number of processors and a higher dimensional
space to search, it is essential to spread the search out, especially in the beginning.
As the search progresses, the radii of the circles become smaller and so the search
can become more local over time. RODDS maintains the DDS method of perturbing
only a fraction of the dimensions. RODDS has a formulae for the rate of decrease of
the radii that does not require calibration for algorithm parameters. Section 6 gives

a precise description of the RODDS algorithm, which is available as open source


code.

Вам также может понравиться