https://github.com/cran/Monomvn
Tip revision: a87c9e13e214bd44021692b6219b730474b24d90 authored by Robert B. Gramacy on 07 March 2010, 00:00:00 UTC
version 1.8-5
version 1.8-5
Tip revision: a87c9e1
ChangeLog
1.8-5 (7 March 2011)
-----
minor chance to regress.lars to correct for cv.lars output
format change in latest lars package
1.8-4 (23 Feb 2011)
-----
Updated rhelp.c file with new my_r_process_events
fixed off-by-one error in Cp method in regress.lars
1.8-3 (23 April 2010)
-----
fixed another error pointed out by Ripley
1.8-2 (21 April 2010)
-----
replaced isinf by R_FINITE as suggested by Brian Ripley
1.8-1 (03 Mar 2010)
-----
now doing n_j <= i + 2 for stabel regression check in
Bmonomvn
fixed rao.s2 bug that was causing a segmentation fault
1.8 (13 Jan 2010)
---
implemented the Griffin & Brown Normal-Gamma prior in the
Park & Casella lasso prior framework
1.7-4 (16 Nov 2009)
-----
added an option to nave no (implicit) intercept in the model
by popular demand
fixed embarassing typo in log_likelihood calculation that
was causing NaNs to be returned for the log posterior
1.7-3 (18 Aug 2009)
-----
rolled out Horseshoe for bmonomvn
fixed a bug in bridge that was causing horseshoe to be run
instead
1.7-2 (2 June 2009)
-----
removed assertion in Blasso::RJup due to check failure for
OSX in bridge example; assertion may be switched back on
when DEBUG = TRUE
1.7-1 (27 May 2009)
-----
added llik_norm hooks to calculate the Bayes Factor
between the Student-t and normal models
initial implemention of Horeshoe shrinkage -- bhs
function available but undocumented
changed to more agressive (larger) thinning when Student-t
errors are being used -- now allow real-vaued thinning
level in bmonomvn. Mixing in nu is improved but the
code is now slower
1.7 (6 April 2009)
---
now calculating the covariance of the mean vector in C
for the predictive covariance calculation in R
implemented Student-t via scale mixtures (Geweke, 1993),
including the pooled-nu version in bmonomvn for
multivariate-t inference
added subset argument to plot.blasso in order to reduce
clutter in tau2i and omega2 printing
now calculating the likelihood for each sample saved
from the posterior in blasso
propogated llik to Ellik calculation in bmonomvn
capm = TRUE now default for bmonomvn, and now allowing
capm = TRUE when method = "lasso"
made RJ = "p" the default
accumumating S != 0 and Si != 0 probabilities in
bmonomvn
added pnz argument to randmvn to allow the specification
of the number of non-zero betas in each round of the
buildup of mu and S, thereby deternining the number of
non-zero entries of S
1.6-1 (3 Dec 2008)
-----
valgrind found some errors and leaks in last version,
now fixed
1.6 (2 Dec 2008)
---
added R matrix to C code, and now calculting the
normalized X considering only those with R=0
added Data Augmentation to complete the monotone
missingness pattern in bmonomvn
added financial returns data set from NYSE and AMEX,
with the market (S&P500) return to be used as a factor
fixed bug in new integrated RJMCMC implementation
forcing the first p columns to be the most observed
when using factor regressions in monomvn; may make
this more general later
added "factor" method to bomonmvn and allow QP
solutions to ignore factors by specifying an integer
argument to bmonomvn rather than TRUE
1.5 (13 Nov 2008)
---
implemented Throughton & Godsill RJ proposals that
integrate out beta -- need to consider integrating
out s2 as in tgp
added mprior argument to allow Binomial(m|M,Mprior)
option when mprior != 0, and Unif over 0,...,M
otherwise
switched back to Ellik.norm for bmonomvn.Rd examples
now tallying lpost in the traces of blasso regressions
within bmonomvn
added plot.monomvn which currently only supports
visualization of the variance of mu and S under the
Bayesian posterior (bmonomvn), and samples from the QP
solutions; see below
added structures for calculating the mean and variance
of mu and S to accomodate the MAP versions of the same
added the ability to sample from the posterior
distribution of the solution to a Quadratic Program
based on (samples) of mu and Sigma
1.4-1 (18 Sept 2008)
-----
Ellik now returns NA if sechol(S1) doesn't work
fixed bug in obs estimator
added sqrt to rmse.muS
now using unbiased cov calculations which has led to
improved performace (drastically in some cases)
added the "factor" method option to do monomvn style
(known) factor based estimates of mu and Sigma
switched back to Ellik.norm for monomvn.Rd examples
1.4 (20 July 2008)
---
removed need to store XtX in blasso explicitly, relying
instead on A and Xtx_diag
added economy option to bmonomvn to save on storage
when running several (thousand) blassos, as would
happen for large ncol(y)
precomputing full Xnorm and X (normalized) in the
Bonomvn module so that it doesn't need to be duplicated
(in smaller chunks) within all Blasso modules
corrected the bn0 summary.blasso calculation to use
the right sampling (-burnin) range and denominator
made regression traces in bmonomvn be stripped down
"blasso" class objects so that the blasso methods like
print and summary can be used
made thinning level dynamic in bmonomvn, depending on
the regression model (i.e., lasso) and RJ
smaller examples throught (b)monomvn for faster R CHECK,
and smaller final blasso example comparing to ML lasso
added parsimonious method for generating random mu and
S within the randmvn function
Ridge regression implemented as a special case in the
blasso function (added bridge function as more direct
interface)
lambda2 now walks under the prior when m=0 under lasso,
but not under ridge; fixed s2 to use Jeffrey's prior
ecaluation in log_posterior
added code to check for monotone pattern in bmonomvn
that will provide the indicator matrix needed for
monotone data augmentation in a later version
added Ellik.norm function to calculate the expected
log likelihood of an estimated normal distribution
relative to the true but don't actually use it
replaced code in the documentation files to use a new
rmse.muS function as a metric for comparison instead
of kl.norm
1.3-1 (19 June 2008)
-----
added cleanup feature to allow interupts within C code
to blasso and bmonomvn
also fixed a bug involving reading in the traces of
bmonomvn regressions
RJ models start in 90% saturate model now, instead of
100% saturated
added the ability to specify starting mu, S and lambda
values for bmonomvn (where appropriate)
now extracting lambda from lars and ridge methods for
use as starting values in bmonomvn
now providing estimats of the variance of the estimates
for S, via the output S.var
added batch option to monomvn function in order to
force a regression for each column (when FALSE)
1.3 (9 June 2008)
---
added calculations of log posterior in blasso
added reversible jump (RJ) variable selection where
proposals come from "full updating" a la Brooks & Ehlers.
This directly effects the Bayesian Lasso -- adding a
new RJ option among others. Of course, this method is
now available to the bmonomvn algorithm as well
added trace plots to plot.blasso for s2 and lambda2
in addition to new model order option "m" when RJ=TRUE
changed lamnda2 default prior setting from (r,d)=(1,1)
to (2,0.1) to avoid singular regressions when RJ is used
with the improper s2 prior with (a,b)=(0,0)
changed the way bmonomvn and blasso take many of its
hierarchical prior arguments
1.2 (8 May 2008)
---
fixed syntax errors in the documentation files and filled
out the INDEX file
Implemented Bayesian Lasso version resulting in a new
function called bmonomvn. A standalone interface to
the Bayesian Lasso is also provided via a blasso
function
added a symmetric option to kl.norm
1.1-4 (11 Nov 2007)
-----
fixed lambda recording bug in regress.ridge
1.1-3 (05 Nov 2007)
-----
added verb=2 and verb=3 arguments for printing each of
the ML regression estimators, and new mu entries and
cols of S
removed verbose LGPL in DESCRIPTION
1.1-2 (31 Oct 2007)
-----
summary.monomvn now calculates the number of zeros in
the covariance matrix (& inv covar matrix) when using the
lasso method -- for summarizing pairwise marginal (and
conditional) independence
a new plot.summary.monomvn function makes histograms of
the numbers of zeros in each column of the covariance
matrix and its inverse
kl.norm now returns Inf with a warning if there is a
truly non-positive definite matrix
increased minimum (m) number of observations in rmono
1.1-1 (21 Aug 2007)
-----
still having problems with the "forward.stagewise"
method -- have emailed Hastie
added ridge regression with a silly heuristic on the lower
lower bound for lambda when p>n
added validation="Cp" for Mallows Cp method for choosing
the best lars model
1.1: (13 Aug 2007)
----
added lars package support for lasso, etc; ridge
regression does not seem to work for the big-p small-n
problem
another special case in addy len(m1)=1, possibly due
to the s22.1 calculation that was modified in the last
version
1.0-3: (03 Aug 2007)
------
added url to DESCRIPTION and each relevant .Rd file
fixed verbosity/validation error in the call of regress
from addy
simplified s22.1 calculation in addy to follow Stambaugh
fixed pls author last name
removed print(full) from rmono
1.0-2: (10 July 2007)
------
added validation argument to allow forcing of LOO
ab=NULL is now the default in rmono.R
cosmetic changes to documentation, particularly for
monomvn
1.0-1: (11 June 2007)
------
error -> stop in posdef.approx