1.2-3
- added support for package 'parallel' (removing support for 'multicore' and older R versions)
- removed '\synopis' from documentation, resulting in some 'hidden' arguments
to be visible now
1.2-2
- fixed bug that prevented subsetting (and cv.GAMBoost) from working
- fixed bug where wrong name would be printed for selected smooth components
1.2-1
- speed improvements for continuous response models
- implemented "criterion='score'" for all types of response
- fixed bug where 'estimPVal' would with only one boosting step
- 'estimPVal' now also works for zero boosting steps
- added auomatic conversion of 'x' and 'x.linear' to class 'matrix'
- creating a copy 'x.linear' is now avoided, if possible
- improved output of 'print' and 'summary' methods
- 'trace=TRUE' now shows the covariate names
1.2
- added function 'estimPVal' for permutation-based p-value estimation
1.1-1
- removed storage of covariate values in return object
- implemented use of score function as selection criterion for linear components
- added support for 'multicore' package for cross-validation
- reduced memory consumption, speed-up for large number of linear components
- added an option for fitting on subsets of observations
- The coefficient matrix 'beta.linear' is for the linear components is
now stored as a sparse matrix, employing package 'Matrix'
1.1
- implemented penalty modification factors and penalty change distribution
via a connection matrix
1.0
- added wrappers (GLMBoost and predict.GLMBoost) for conveniently fitting generalized linear
models, i.e., without smooth components
- fixed "zero boosting steps" corner case:
- GAMBoost/GLMBoost can now fit with "stepno=0", the trace, deviance, AIC, and BIC
result vectors have an additional element for boosting step zero, and the elements
of 'beta'/'beta.linear' for the latter are no longer equal to zero, but contain the
results from one estimation step for the mandatory covariates, i.e., in the simplest
case, an intercept-only model is fitted at zero boosting steps.
- cv.GAMBoost/cv.GLMBoost can return an optimum at zero steps, and alos the 'criterion'
and 'se' elements of the results have an additonal element for boosting step zero
- implemented parallel evaluation on a compute cluster for cross-validation
0.9-4
- general performance improvements, especially for componentwise ridge boosting, i.e. boosting
for covariates with linear influence, and there espacially for binary response models
- fixed bug in formula for traditional AIC in the Gaussian response case
0.9-3
- fixed use of weights in cv.GAMBoost
- added flexible p value cutoff for prediction and calculation of prediction error
in cv.GAMBoost
0.9-2
- fixed a problem where predict.GAMBoost would not work with only linear predictors
(thanks to Ravi Varadhan for pointing this out)
- implemented penalty of difference '0' as absolute penalty on coefficients
- 'pdiff, specifying the penalty difference can be a vector now, thus allowing for enforcement
of several types of smoothness simultanoeusly'
- fixed bug that prevented criterion from being save for AIC-optimGAMBoostPenalty
- optimGAMBoostPenalty now also stores the selection criterion in the GAMBoost object returned
0.9-1
* initial public release