cocons 0.1.4
Enhances
GetNeg2loglikelihood,
GetNeg2loglikelihoodTaper, and
GetNeg2loglikelihoodTaperProfile now much faster (~35%
faster)
cocoOptim:
- new
REML estimation for dense coco
type.
- now cpp functions
cov_rns and
cov_rns_taper include as special cases the well-known
shapes when nu = 0.5, 1.5, and 2.5, yielding computational
speed-ups.
- now
getCIs display names of covariate instead of index
of Design Matrix.
- now
.cocons.check.info also checks for smoothness model
and smooth.limits.
Changes
coco now accepts data.frame for locs
argument, which is then converted to matrix.
getTrend now is called getSpatMean.
cocoPredict renamed output spatial mean vectors:
trend is now called systematic, while
mean is called stochastic.
cocoOptim :
- reordering of arguments.
mle estimation method now called ml.
pmle estimation method now called
pml.
getHessian :
mle to ml and pmle to
pml.
getCIs now alpha argument reflects
confidence level instead of 1-confidence level.
Fixes
- fix a bug for plotOptimInfo when handling
pml or
reml objects.
cocons 0.1.3
- Vignette updates, watermark removed
- Documentation polished
- New
holes_bm dataset with independent realizations and
spatial trend
Enhances
- added a NEWS.md file with version updates / modifications / enhances
/ etc
- automatized
delta for method plot for coco
class.
- Better visualization for
plot(cocoOptim object, type = "ellipse")
coco:
- now it is not necessary to provide all models for each source of
nonstationarity. Those not specified will be set to those referenced
with a stationary 0-mean model (i.e. tilt = 0 , aniso = 0, nugget = -Inf
(because of log-parameterization)). If ‘smooth’ is not specified, then
it is set to 0.5.
cocoOptim:
- “auto” option for
ncores argument for
cocoOptim, providing a convenient number of threads based
on the number of parameters to estimate, available threads, and settings
of the LBFGSB routine
- “safe” argument, which prevents crashes due to ill-posed covariance
matrices (Choelsky factorization error)
.cocons.check.convergence now checks and reports at
which iteration ill-posed covariance matrices have been found during the
optimization.
- now “pmle” works with multiple independent realizations for coco
types
dense and sparse
- safer parallel handling
getHessian more memory efficient
cocoSim:
- if provided a fitted coco object, then
pars argument
can be NULL (default), and
coco.object@output$par is used instead (and also
type is set to diff).
- more memory efficient
- safer parallel handling for
getHessian
- polishing of neg2loglikelihood functions, leading to more efficient
code
- new and more polished internal functions to assess the validity of
arguments (
stopifnot() instead of
if() stop())
getCondNumber optimized
- small improvement over cpp functions
Changes
- switched
getPen as an internal function
- method
plot for coco objects shows
rotation angle of the kernel w.r.t x-axis
GetSpateffects now provides angle w.r.t to x-axis
- method “summary” for
coco objecets (former “print”
method)
- more proper naming of objects inside functions
- renaming of “cat.vars” to “skip.scale” + associated checks and
optimization
getCondNumber removed, which can be replaced with
kappa function from base R
(i.e. kappa(getCovMatrix(coco.object),exact = TRUE))
cocons 0.1.2
- improve overall help files
- add examples for
coco, cocoOptim,
cocoPredict, and cocoSim
- update Vignette
- some bug fixes and overall code polishing
smooth_limits from info from the coco
function is now called smooth.limits to match the style of
other arguments (no backward compatibility)
- new names for C++ sparse covariance functions
- add warnings of non-convergence of the LBFGSB for
cocoOptim
- bug fixes for coco “methods”
- less redundant code