2.4-9 (01 April 2013)
-----
Changed printed tables to 15 digits to comply with new
read.table checks
removed zzz.R and moved LICENSE to License
global exp2d and data call made fancy to pass checks
2.4-8 (13 Oct 2012)
-----
fixed bimprov/improv bug overlooked in NULL issue
updates
removeed bessel_k.c and assocated functions; now using
built-in bessel_k_ex function which didn't exist before
2.4-7 (18 April 2012)
-----
fixed NULL issue in lh_sample
2.4-6 (14 April 2012)
-----
Added check that E divides T-B
Corrected bug in slice grid causing overflow
Corrected NULL problems pointed out by BR
2.4-5 (30 January 2012)
-----
fixes to vigneetes after NAMESPACE change, and removing
direct references to stdout and stderr
2.4-4 (23 January 2012)
-----
Added NAMESPACE file
2.4-3 (18 December 2011)
-----
Checked for assert problems with NDEBUG and updated R
version (2.14) and email addresses
2.4-2 (30 March 2011)
-----
added dp.sim to tgp.default.params so that MVN proposal
covariance matrix can be specified by the user in GP-SIM
models
2.4-1 (23 Feb 2011)
-----
Changed my_r_process_events to always ProcessEvents and
check for interrupts (every second)
Made new sobol_indices function to clean up tgp::Sens, and
using shape to drive categorical calculations rather than
bmax
minor edit to tgp2 doc to correct fried.bool description, and
edit to fried.bool to fix fried.bool(1) problem
2.4 (20 Sept 2010)
---
added sim (single index model) rank 1 correlation structure
overlooked fixing nugget in treed split proposals; now
fixed
2.3-4 (19 July 2010)
-----
allowed the entire $tree component from tgp.default.params
to be written over by the tree argument to b* functions
fixed a bug in the linear initialization of correlation
functions (for non-LLM models)
2.3-3 (23 April 2010)
-----
fixed another bug pointed out by Ripley
2.3-2 (21 April 2010)
-----
replace isinf by R_FINITE as suggested by Brian Ripley
2.3-1 (8 Feb 2010)
-----
Added reference to second JSS paper
2.3 (23 Dec 2009)
---
final version of the tgp2 vignette, accepted at JSS
removed mrtgp documentation
2.2-4 (20 Nov 2009)
-----
added a check for a full rank design matrix in the
pre-processing of inputs
corrected some syntax errors in .Rd files caught by
new checks on CRAN
changed sens warning to check for nn > 0 instead of null
XX which was causing a warning to be printed every time
regardless of whether or not there was a violation
2.2-3 (22 Jul 2009)
-----
added nug.p=0 option to cause the nugget to stay fixed
at the params$gd[1] value. bcart, btlm, and blm now
automatically use nug.p=0 since the nugget is not
identified in these models
2.2-2 (27 May 2009)
-----
caught a signifigant typo/bug in Xsplit pre-processing
in R which was causing the bounding rectangle to be
computed improperly
removed combinat dependency
2.2-1 (20 Feb 2009)
-----
added Tree::Clear() to rejected prunes in Tree::Prune()
tgp.default.params and tgp.check.params now allow the min
partition size to depend on basemax rather than d
the bounding rectangle and the valid tree-split locations
are now governed by an Xsplit automatically generated in
the tgp R function, and gathered from the tgp object in
predict.tgp
2.2 (16 Jan 2009)
---
added lower bound for s2_g0, and added DEFAULT CPP macros
to help prevent assertions which would have backed out of
bad parameterizations anyways. This more liberal approach
essentially prevents assertions in the no-noise regression
case
now allow basemax != d for constant mean functions
added categorical LHS for sensitivity analyses
finished draft of second vignette, to be submitted to JSS
2.1-6 (25 Nov 2008)
-----
caught important bug in b0-prior pre-processing in R that
was causing the emperical Bayes prior to be used when the
b0 hierarchical one is requested
2.1-5 (24 Oct 2008)
-----
added check for >= d+1 unique rows of X in Tree::Singular
changed how d an nug are printed in the progress meter for
mrtgp
added par(mfrow=c(1,1)) to default 1-d mrtgp plots
2.1-4 (17 Oct 2008)
-----
made it possible to cap the number of input locations which
are ranked by the expected improvement statistic by
allowing the improv argument to b* functions to take a
2-vector input. Also no longer ranks points with an
improvement statistic of zero
2.1-3 (7 Oct 2008)
-----
added check for >1 unique Z in Tree::Singular
allow (undocumented) specification of minpart through the
ellipses (...) argument to the b* functions
m0r1=TRUE is now the default for all b* functions, the Rnw
documentation has been updated (but the pdfs have not been
regenerated in the current version)
2.1-2 (17 April 2008)
-----
introduced params$Wi and made it so that params${b,Wi} set
both the starting an prior values for the corresponding
parameters.
added bprior="bmznot" so that tau2 can be fixed at its
starting value
2.1-1 (4 April 2008)
-----
separaing tcounts and cum_tcounts and resetting tcounts to
the average of cum_tcounts at beginning of each round
fixed bug in returning of tcounts back to R by encuring that
its->UpdateProbs() is always called after the last of R
rounds whenever its->Numit() > 1
return ess information at each inverse temperature back to
R for storage in the tgp-class object
2.1 (22 March 2008)
---
moved CHANGES to ChangeLog for the new CRAN packages
page
cleaned up tgp vignette into one .Rnw file including
several .iRnw files
added linear=FALSE to akima interp commands
stopped printing both separate ess and essN in IT method
added new itemps.barplot function for visualizing the number
of visits to each inverse temperature in the ST-MCMC chain
now tcounts (observation counts) for IT are accumulated
across repeats (R>=2) -- not sure if this is a good idea
now recording tree acceptance rates (Grow, Prune Change
and Swap accepts over total)
propogated splitmin to up to the R interface, and added
a basemax paramter also specifiable in the R inferface --
the parameters allow control of which X colums can
be partitioned upon (by the tree) and which appear in the
base (GP) model
added heights="map" argument to tgp.trees for plotting only
the tree corresponding to the maximum a' posteriori tree
wrote two sections in the new vignette (tgp2): one on
splitmin and basemax (i.e., categorical inputs) and one
on importance tempering
2.0-4 (23 Jan 2008)
-----
made a change suggested by Andrea Spano to make the
partition function more robust when new X values
(on a different range from the original data) are
provided
2.0-3 (27 Dec 2007)
-----
commented the moving average code (for sens) and made
it more efficient
2.0-2 (12 Dec 2007)
-----
allow mode in the sens function to be be within the
bounds of rect -- i.e., stop checking for positive mode
fixed dyslexic error in man page for sens describing
total effect indices
fixed missing braces in some .Rd files
shortened to LGPL in DESCRIPTION
2.0-1 (1 Oct 2007)
-----
Memory savings: preds->ZZ and preds->Zp only allocated when
pred.n=TRUE, preds->w and preds->itemp only allocated when
ST or IT or IS
corrected default.itemps() example barplot and changed
k.min to 0.1 to agree with IT paper
fixed zcov bug in tgp.postprocess
2.0 (21 Sept 2007)
---
consolidated R functions from individual files into those
grouped by task
made gridlen argument a 2-vector so the grid size can be
specified in both x and y
new economy argument to tree and Base duplicators so that
the entire set of covariance matrices is not copied unless
necessary
Matt Taddy officially joins the list of authors, adding
multi-resolution tgp and sensitivity analysis
importance tempering, with importance sampling and
simulated tempering support
Latin hypercube sampling with respect to a Beta
distribution
moved most of the reading of parameters out of the Tgp
constructor and into Tgp::Init so that tgp_cleanup() can
safely delete the Tgp module even it it wasn't successfully
initialized
ordered multiple improv statistics
predictive covariances now calculated instead of variances
1.2-7 (19 Aug 2007)
-----
tgp.partition 1-d data handling made more robust
removed assertion on w that tends to fail for unknown
reasons
1.2-6 (17 June 2007)
-----
added number of leaves to traces
fixed CART bug in Gp::predict that was already handled in
v1.3, thanks to Taddy, but was bad in the CRAN version
1.2-5 (29 May 2007)
-----
allowing btgp to do a linburn
$parts is a now a matrix, and $linburn now {TRUE,FALSE}
$tree and $hier removed from tgp-class output, and
$trace correctly removed when input trace=FALSE
got rid of some hanging code fragments in documentation
of btgp and dopt.gp
1.2-4 (07 May 2007)
-----
fixed a leak when krige=FALSE appearing in Tree::Predict
fixed sub-vignette cutting for preds and traces
corrected documentation of nug.p in tgp.default.params()
added verbosity and iter arguments to dopt and thus
dopt.gp and tgp.design
1.2-3 (09 Apr 2007)
-----
updates to the vignette for the JSS publication
check.matrix modified to allow one-column matrix Z
arguments which have column-name information
tgp.trees calculation of rows and columns corrected for
situations when there is a height 1 tree
1.2-2 (19 Dec 2006)
-----
slice.interp bugs fixed (is.null(NULL) and plotting of
axis labels) and tick marks added
added R argument to predict.tgp and changed the krige
argument to be called MAP
added new krige argument to b* and predict.tgp to allow
the user to control whether Z*k* kriging samples are to
be obtained (possibly saving on memory)
bug involving the order of trees read from files in the
tgp.get.trees function is fixed; as a result the $trees
output field in "tgp"-class objects is listed by height
and so may contain NULL entries
documented plot.tgp and predict.tgp with the method
markup
1.2-1 (03 Dec 2006)
-----
Added "cart-" option to bprior in order to allow a
CART style mean rather than a LM mean
1.2 (21 Nov 2006)
---
Now allowing BTE[2]-BTE[1] % BTE[3] != 0
Added predict.tgp and necessary functionality to save
and load in MAP tgp model for later prediction
Corrected par calculation in tgp.trees for non-default
which argument when there is an unprinted height 1 tree
Traces not recorded for linarea, parts, and ego, when
they are off or fixed to unchanging settings
Fixed EGO calculations; now calling them "improv", made
them thread-safe, and added improv samples to traces
Also now renamed ds2x to Ds2x, and made Ds2x (i.e., ALC)
calculations thread-safe
Ellipses to b* function allow other params arguments to
be passed to replace defaults in tgp.default.params()
Added traces of hierarchical parameters, and all
predictive summaries (including kriging means and vars)
Initial implementation of Importance Tempering idea
Calculating variances of predictive distribution, and
saving kriging means and variances
Fixed numPrunable (P+1) calculation in grow attempts
EGO calculations made thread-safe
1.1-11 (20 Sept 2006)
------
Bad compilation flags caused bad bug in log_HierPrior
to go unnoticed until assertion failure on Windows
1.1-10 (19 Sept 2006)
------
Fixed big compute_ego() bug
Tweaks to RNG seeds so that they give same sequence
across platforms (tested on OSX/PB & Linux/Xeons)
Allow X=NULL in dopt.gp()
Removal of constants LOG_2_PI covered by R internal
M_LOG_SQRT_2PI
Added mvnpdf_log_dup and wishpdf_log functions
Corrected full log-pdf function for tree prior, and
added log-pdf of W and B0
1.1-9 (03 Sept 2006)
-----
Many comments added in predict[_linear].c and exp_sep.cc
Fixed ds2x (predict_linear) memory leak
Fixed Matern bug -- now using correct pwr=1 distance
computation; added comments to gen_covar.c
Compilation warnings addressed in update of compat-
ibility with C/C++ adaptive sampling (as) code
Now getting BestPartitions from Tgp instead of at the
end of Model::Rounds
1.1-8 (29 Aug 2006)
-----
Added trace capability for Matern corr family
Fixed bug that num_consumed is not updated on final
NULL read from tlist
Using new log_bessel_k() function to bypass allocs of
floor(nu)+1 double each time bessel_k() was called
1.1-7 (23 Aug 2006)
-----
Added the ability to get traces of all parameters,
except the hierarchical ones
Fixed bug to undo.m0r1 so that quantile differences
don't have the mean subtracted out twice
Changes to allow "..." arguments to mapT, and plotting
of 2-d slices of input locations
Minor fix to do params->Print() within model->Print().
Begin addition of mr_tgp, by Matt Taddy. When these
changes are complete, we should be moving to version
2.0
1.1-6 (26 Jun 2006)
-----
Added Latin Hypercube design function
Added "mapT" function for stand-alone plotting of MAP
partitions via projection and slice. Fixed bug in
"tgp.plot.slice" in the process
Also now calling "tgp.get.partitions" by the shorter
name "partition"
1.1-5 (27 May 2006)
-----
Fixed memory leak with RNG state in predict_consumer
Solved -DPARALLEL adaptive sampling mystery
exp2d.rand now returns more information about responses
including ZZ, Ztrue, and ZZtrue
1.1-4 (22 May 2006)
-----
Instructions for compiling with Pthreads.
Got rid of duplicate printing of d[a,b] in tgp header.
Added verb arguments to specify how verbose the
R-console printing should be be.
1.1-3: (16 Apr 2006)
------
Necessary changes in order to get read_ctrlfile to work
for the command-line (non-R) interface to tgp, and other
minor changes.
Noticed bug for bcart and b0tau options due to bad T-
matrix initialization -- fixed.
1.1-2: (04 Mar 2006)
------
1. Beta version of Matern correlation function working
(abiet a bit shakily)
2. Tgp class and cleanup function now implemented in
order to let tgp be interrupted, and then clean up to
avoid memory leaks "on.exit()"
1.1-1: (26 Feb 2006)
------
Minor edits to documentation.
1.1: (22 Feb 2006)
----
1. Modularization overhall of C/C++ code for passing
to Matt Taddy, who will be responsible for such future
additions as Matern correlation functions, and multi-
resolution GP implementations.
2. Plotting enhancements, including 1-d projections
and (now default) loess interpolation for 2-d surfaces
and projections, plotting of ALC & EGO statistics,
and ability to control plot layout to show mean surface,
error surface, or both.
3. URL/Email now to UCSC/AMS address.
1.0-2: (Dec 2005)
------
1. Wrote generic random deviate function which can use
unif_rand(), randomkit, and erand48.
2. Fixed up some documentation bugs and typos.
3. system command in model.cc no longer calls "rm" which
doesn't work on Windows.
1.0-1: (Dec 2005)
------
1. No longer using erand48; using unif_rand() from R for
Windows compatibility.