![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
spsann - optimization of sample patterns using spatial simulated annealing |
VerfasserIn |
Alessandro Samuel-Rosa, Gerard Heuvelink, Gustavo Vasques, Lúcia Anjos |
Konferenz |
EGU General Assembly 2015
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 17 (2015) |
Datensatznummer |
250108049
|
Publikation (Nr.) |
EGU/EGU2015-7780.pdf |
|
|
|
Zusammenfassung |
There are many algorithms and computer programs to optimize sample patterns, some private
and others publicly available. A few have only been presented in scientific articles and text
books. This dispersion and somewhat poor availability is holds back to their wider adoption
and further development.
We introduce spsann, a new R-package for the optimization of sample patterns using
spatial simulated annealing. R is the most popular environment for data processing and
analysis. Spatial simulated annealing is a well known method with widespread use to solve
optimization problems in the soil and geo-sciences. This is mainly due to its robustness
against local optima and easiness of implementation.
spsann offers many optimizing criteria for sampling for variogram estimation
(number of points or point-pairs per lag distance class - PPL), trend estimation
(association/correlation and marginal distribution of the covariates - ACDC), and spatial
interpolation (mean squared shortest distance - MSSD). spsann also includes the mean or
maximum universal kriging variance (MUKV) as an optimizing criterion, which is
used when the model of spatial variation is known. PPL, ACDC and MSSD were
combined (PAN) for sampling when we are ignorant about the model of spatial
variation. spsann solves this multi-objective optimization problem scaling the objective
function values using their maximum absolute value or the mean value computed
over 1000 random samples. Scaled values are aggregated using the weighted sum
method.
A graphical display allows to follow how the sample pattern is being perturbed
during the optimization, as well as the evolution of its energy state. It is possible to
start perturbing many points and exponentially reduce the number of perturbed
points. The maximum perturbation distance reduces linearly with the number of
iterations. The acceptance probability also reduces exponentially with the number of
iterations.
R is memory hungry and spatial simulated annealing is a computationally intensive
method. As such, many strategies were used to reduce the computation time and memory
usage: a) bottlenecks were implemented in C++, b) a finite set of candidate locations is used
for perturbing the sample points, and c) data matrices are computed only once and then
updated at each iteration instead of being recomputed.
spsann is available at GitHub under a licence GLP Version 2.0 and will be further
developed to: a) allow the use of a cost surface, b) implement other sensitive parts of the
source code in C++, c) implement other optimizing criteria, d) allow to add or delete points
to/from an existing point pattern. |
|
|
|
|
|