Versions 0.01 - 0.19
This is an experimental package for exploring various empirical Bayes
problems usually involving Kiefer-Wolfowitz non parametric ML estimation
of various mixture problems. It supplants my earlier MeddeR package that
had some similar capabilities, but employed an interface to Mosek via Matlab.
Note that there is a .Rprofile file that may be needed to specify the
license file for mosek. This may be needed to get mosek() to agree to do something.
At some point Rmosek became capable of doing multicore things and simulations
became much more convenient using foreach().
Version 0.20
1. Added the tuning parameter rtol to KWDual and friends to control the
convergence tolerance. The default value of 1e-6 is the same as Mosek
but in some cases tightening it to say 1e-10 can produce a better solution.
There is a demo called "tannenbaum" that illustrates this.
Version 0.21
1. Fixed a bug in GVmix that was caused by my neglect of a Jacobian (mea
culpa!). This shouldn't affect simulation experience simv[123].R since it
was just a multiplicative factor in the loglikelihood. But this needs to
be checked.
Version 0.22
1. Added Bmix function for Binomial mixtures and a demo of this using the
Beckett and Diaconis tack data taken (shamelessly) from the DPpackage.
2. Added rtol parameter to the call for WGLVmix and WGVmix.
3. Some further cleaning up of bball: name matching and removal of observations on
players with fewer than 3 half seasons.
Version 0.23
1. Fixed the documentation for tacks data.
2. added option to collapse binomial data into cell counts in Bmix.
Version 0.26
1. For hist = TRUE option in GLmix the binning is now done with equally
spaced bins on the full support of the data. The parameter m can be used
to choose the bin width. And v can be given to separately control grid for
fit.
Version 0.27
1. Changed KWDual to return weighted log likelihood. This was ok as it
was when the weights were always 1/n but for the histogram binning, it needed
to be changed. Now it returns the average loglikelihood so for example it
GLmix it gets multiplied by n to get the usual loglikelihood value.
Version 0.28
1. Added TLmix to do mixtures of Student t's with known df. Note that the
equal quantile spacing of the v's is experimental.
Version 0.30
1. Cleanup of some issues in the medde.Rd file to prepare a version for CRAN.
Version 0.31
1. Reduced size of example in medde.Rd and moved other examples that ran
demos out of examples and into details section.
Version 0.32
1. Reduced size of examples again in medde.Rd and added SystemRequirements line to
Description file.
Version 0.35
1. Bug in TLmix that messed up the default v grid. (Damn missing minus signs!)
2. Changed the default grid in TLmix back to equally spaced from quantile
based. This probably needs further investigation.
3. Added pv to return in GLVmix.R and to WGLVmix.R
4. Added 2012 data to bball
5. Added g to the returned list from GVmix and WGVmix
6. Added TLVmix and WTLVmix to do Gaussian mixture estimation with Student
t formulation of the likelihood.
7. Added Gompertzmix and Weibullmix and the medflies data set.
Version 0.37
1. Changed the definition of g output by GLmix to agree with its man page.
If we ever get around to unifying all these functions this needs to be checked
carefully, there are various issues especially when there are weights
involved.
Version 0.38
1. Fixed a bug in the medde function for the monotonized Bayes rule
estimator.
Version 0.41
1. Removed GLVmix and WGLVmix since they are supplanted by TLVmix and
WTLVmix.
2. Added a predict.GLmix function as an illustration for how to do prediction
from the posterior for Lp norm loss with p in {0,1,2}.
3. Added a sigma argument to GLmix so that one can specify non standard
Gaussian noise.
4. Added a demo for prediction to illustrate (3.).
5. Added check for zero indices in the Loss = 1 predict.GLmix function