Changes to Earth Package ------------------------ 3.2-7 Jan 28, 2014 Tweaks to standalone version of earth for clean compiles with certain 64 GCC builds. plot.evimp now works with intercept-only models. Other minor tweaks to plot.evimp. Clerical changes to satisfy recent CRAN checks. 3.2-6 Apr 15, 2013 Added format.earth style="C". Included the leaps package files into the earth package, to prevent warnings from CRAN check about earth's use of internal leaps code. Removed "Depends: leaps" from earth's DESCRIPTION file. Lin deps discovered by leaps are now slightly better handled. Small changes to earth.c and earth.h for stand-alone builds. Added sections to the FAQ on using earth for binary and categorical responses. Added a section to the vignette on building a model based on cross validation results. Removed trace=4 from test.earth.R so a non-problem is no longer reported due to minor numerical differences across architectures. Reduced the number of digits printed in some of earth's trace messages. This makes trace results more consistent across different architectures (the ls digits are often in the numerical noise floor). We now print a max of 20 columns of a matrix when tracing. The legend.text arg of plot.earth.models now works correctly. 3.2-3 Apr 27, 2012 Incorporated Glen Luckjiff's patch to fix predict.earth(type="terms") failure when bx had just two columns. We no longer call fflush when in an R environment. The tests/test.earth.R code now uses the tree rather than the ozone1 data, in an effort to get test consistency across different architectures. Added .Rinstignore to drop inst/docs/figs files (prevents note in CRAN build). 3.2-2 Mar 14, 2012 We no longer pass R_NilValue from the R code to the C code. Incorporated Olaf Mersmann's patch for POS_INF with the INTEL_COMPILER. Code touchups for Microsoft Visual Studio 2010. 3.2-1 Jul 19, 2011 Documentation touchups. 3.2-0 Jun 19, 2011 The PDF file is now a standard vignette (but with no executable code). The "Importances" printed by print.earth are now a maximum of option("width") wide. We no longer warn "effective number of GCV parameters >= number of cases" but instead set such GCVs to Inf, so GCVs are now always non-decreasing with the number of terms. Documented as FAQ 12.11. The earth C routines now print memory allocations if trace=1.5. The pnTrace arg to ForwardPassR is now a double and called pTrace (allowing trace=1.5). The default nk is now clipped at 200. This limits the amount of memory needed in the forward pass for large x matrices. Numerical limits will probably kick in before 200 are reached anyway. We now force Use.beta.cache=FALSE with a message if the beta cache would be more than 3 GB. summary.earth now warns if illegal arguments are passed. Documentation touched up. 3.1-1 Jun 15, 2011 Documentation emendations. We now issue a warning if user uses a CV arg to plot.earth but no CV data to plot. The CV summary print no longer prints the response name for single response models. 3.1-0 Jun 14, 2011 print.evimp now formats the results a little differently (more clearly). "Reached min GRSq" and similar trace messages now also print the number of terms. The min auto-calculated ylim in model selection plots is now clamped at -1. Added the col.residuals, col.pch.max.oof.rsq, and col.pch.cv.rsq args to plot.earth. Cross validation with wp no longer gives an error message. Revised the documentation. 3.0-0 Jun 10, 2011 Created the vignette earth-notes.pdf which collects and extends documentation that was previously in the help files. Revamped cross validation. New argument "ncross" to earth. Extended plot.earth, mostly for cross validation. The default for evimp's "sqrt." argument is now TRUE. plot.earth now uses lowess rather than loess (loess tends to issue ugly warnings). Removed earth's Print.pruning.pass argument. Earth now uses less memory if passed an x matrix of doubles. Earth's memory report (with trace>=4) was not accurate in all cases and has been removed. Tweaked the Author: entry in DESCRIPTION for better results from citation(). Other minor touchups. 2.6-2 Apr 25, 2011 Moved plotmo.methods.earth.R to here so the plotmo package is independent of earth. Modified test scripts to conform to R 2.13.0's way of printing numbers. 2.6-1 Apr 13, 2011 Fixed an issue of dependency on the plotmo package. 2.6-0 Apr 12, 2011 Moved plotmo to the plotmo package, and added "depends plotmo" to earth's DESCRIPTION Added the Exhaustive.tol arg as a work-around for leaps.exhaustive error code -999 (Thanks to David Marra for help tracking this down.) The "Exhaustive pruning" pacifier is now printed based on the number of subsets Warnings from leaps routines are now treated as errors (because ret val is bad with warning) Replaced n=2 and similar in calls to eval with an explicit environment The decision to use eval.subsets.xtx is now more automatic and less obtrusive The response y is now named in certain situations where it used to be unnamed Removed annoying warning "specified nprune 20 is greater than the number of model terms" Some code tidying 2.5-1 Mar 25 2011 Modified handling of illegal arguments 2.5-0 Mar 24 2011 Changes to plotmo: The arguments were reordered (moved the most important arguments first). The ycolumn arg is now called nresponse and handled more consistently. The degree2="all" setting is now replaced by all2=TRUE, likewise for degree1. The default for the type argument is now object dependent. The cex parameter is now supported, and text is a little larger for most plots. The degree2 graph titles now have a space for readability. There is some minor backwards incompatibility in the minor arguments. The plotmo graphs are now ordered on the variable order, for all model classes. Plotmo now handles a wider range of object classes. Factor predictors are now plotted in degree2 plots. Miscellaneous other touchups. 2.4-8 Mar 05 2011 plotmo now handles vector colors better plotmo is now better at plotting lda and qda models Added the jitter.response argument to plotmo Moved plotmo method functions into a new file, plotmo.methods.R. 2.4-7 Feb 02 2011 Intercept-only models are now better supported, esp. for earth-glm models. Martin Renner and Keith Woolner provided the impetus for this change. Legend arguments are now handled better in plot.earth.models. A vector legend.text is now handled correctly and you can use values like "topleft" in legend.pos (thanks to Gavin Simpson at ECRC). The elements of cv.list in earth's returned value are now named, and plot.earth.models(model$cv.list) thus has better legends. Cumulative distrib plots in plot.earth.models now use the jitter arg, and plot.earth's (pointless) jitter argument has been removed. A few code and doc touchups were also made. 2.4-6 Jan 26 2011 Made the following changes to plotmo: We now print the fixed grid values unless trace < 0. The degree1 and degree2 arguments now have an "all" option. We now have better support for rpart objects The cex arg is now also used for response values (with cex.response!=0). Added cex.lab argument. For factor predictors we now print the factor levels vertically along the x axis. Fixed crossed type.gcv and type.rss args in plot.evimp. Fixed residuals.earth to handle vector responses correctly. (Thanks to Gavin Simpson at UCL for providing fixes for the above bugs.) plot.evimp now doesn't use on.exit(par(no.readonly=TRUE)), so we now allow subsequent plots on the same page with mfrow != c(1,1). Did the usual document touchups. Minor cleanups for R check (partial arg matches). 2.4-5 Oct 30 2010 Plotmo now better handles the "data" argument for earth.formula models. (Thanks to Keith Woolner for spotting that this was previously incorrect.) 2.4-4 Oct 6 2010 NAs are now allowed in the data passed to predict.earth. The predicted value will be NA unless the NAs are in variables that are unused in the earth model. 2.4-3 Sep 29 2010 For certain data we were seeing duplicated term names for cuts that are very close to each other. To avoid this you can now use options(digits) to increase the number of significant digits used when forming term names. (Thanks to Keith Woolner for spotting this.) 2.4-2 Aug 28 2010 We now give the correct error message for attempted cross validation of paired binomial responses (which is not yet supported). Labels on the largest residuals in plot.earth graphs are now slightly better positioned (using plotrix::thigmophobe.labels). plotd now handles qda objects in the same manner as lda objects Other doc touchups. 2.4-0 Nov 8 2009 This version of earth will build models SLIGHTLY DIFFERENT FROM PRIOR VERSIONS, because I fixed a bug in the calculation of minspan and endspan. Use minspan=-1 for backwards compatibility. (Thanks to Gints Jekabsons for spotting this.) 2.3-5 Oct 7 2009 Fixed bug where glm.control args were ignored under certain circumstances (Thanks to Jérôme Guélat for letting me know about this.) 2.3-4 Sep 21 2009 . Fixed bug where predict.earth of a earth-glm model under R 2.9.0 and higher incorrectly gave the message: Object of type 'closure' is not subsettable (Thanks to Thomas Brockmeier for help tracking this down.) . plotd now correctly ignores the vline if vline.col=0 . Doc touchups 2.3-2 Mar 23 2009 Extended plotd Doc touchups 2.3-1 Feb 26 2009 Doc touchups Added leaps::: prefix needed for new version of the leaps package Added style="max" to summary.earth Added labels param to plotd Reordered some of earth's arguments 2.3-0 Feb 18 2009 Added the plotd function Removed some restrictions on type="class" in predict.earth Added sqrt. argument to evimp Added more grid lines to cumul density plot in plot.earth Added a help page section on interpreting the graphs in plot.earth Miscellaneous other touchups to code and docs 2.2-3 Feb 2 2009 Added levels to the return value Added type=class and thresh arguments to predict.earth Thanks to Max Kuhn for suggesting these improvements 2.2-2 Jan 30 2009 Doc touchups Cross validation MaxErr is now signed 2.2-1 Jan 22 2009 Added cross validation i.e. the nfold and stratify parameters. We now scale y before the forward pass, for better numeric stability in the forward pass with very big or very small y's. For the old behaviour, set earth's new argument scale.y=FALSE. Added get.pairs.bagEarth so plotmo prints degree2 plots for caret:bagEarth models. Changes internal to earth.c: High values of trace argument are treated differently ServiceR (to allow interrupts) is called more consistently for large datasets Delta RSS handling is simplified Changed some var names for consistency The last two were a byproduct of experimental changes to earth that were not included in this release. Changed documentation to American English. 2.1-2 Nov 18 2008 Touched up evimp help page. 2.1-0 Nov 15 2008 plotmo now has better support for factors and for glm models na.action (always na.fail) is now handled as documented in earth.formula Added style="bf" to format.earth and summary.earth Fixed a few minor bugs and touched up documentation for evimp 2.0-6 Oct 30 2008 You can now pass only the needed subset of columns to predict.earth Added plot.evimp Removed spurious warning "Need as many rows as columns" 2.0-5 Jul 14 2008 Touched up documentation for format.lm. 2.0-4 June 22 2008 Touched up code and documentation for a zero thresh value. 2.0-3 June 22 2008 Zero values are now allowed for earth's "thresh" parameter (previously if you used thresh=0, thresh was clamped internally to 1e-10). Also, if thresh=0, the MAX_GRSQ forward pass condition is ignored. The idea is to get as close to a big nk as possible Changed "valid.names" argument of format.earth and format.lm to "colon.char" which achieves the same end more simply. 2.0-2 June 15 2008 Added column names to results of mars.to.earth, allows use of evimp. Added valid.names argument to format.earth and format.lm. 2.0-1 June 10 2008 Added "namesx" and "first" arguments to the "allowed" function. evimp() for a scalar x now returns a matrix (I added a missing drop=FALSE). 2.0-0 June 07 2008 Added support for glms and factors (but plotmo does not yet support factors). Added variable importance function "evimp". Added response weights argument "wp" to earth. Output of summary.earth has changed to better deal with multiple response models, see the "style" argument. Added namesx and namesx.org to earth's return value. 1.3-2 Mar 29 2008 Fixed two bad multiple response bugs: a) for multiple reponse models, earth calculated the wrong null RSS and therefore the wrong RSq and GRSq for the sub-models. (The total RSq and GRSq were correct.) b) the wrong betas were used when pruning multiple response models. Also fixed a bug where summary.earth printed the wrong number of cases for multiple response models. 1.3-1 Mar 22 2008 "update.earth" now has a "ponly" argument, to force pruning only. Because of this change, the "ppenalty" argument to earth is no longer needed and has been removed. Revisited text of warnings after I was bamboozled by one of my own warnings. Tweaked legend positioning in plot.earth.models. 1.3-0 Mar 18 2008 Default minspan is now 0 (was 1) for compatibility with mda:mars and Friedman's MARS paper (I've flip flopped on this one). This means that models built with the default args will be little different to before. Earth's peak memory is now about 40% less. Big models are now more responsive to ^C. For multiple response models, we now print response names in most places instead of just "Response N". Removed get.nterms.per.degree and get.nused.preds.per.subset from NAMESPACE and from help pages, to simplify user interface. Fixed some niggling document issues and extended the FAQ. 1.2-2 Jan 2008 print.summary.earth now prints the call even for x,y interface to earth plotmo now accepts x matrices without column names Tweaked FindKnot and OrthogResiduals for speed Removed a few shadowed variables in earth.c after running gcc -Wshadow Clarified some paras in earth.Rd, reduced page width for better html display 1.2-1 Fixed a newvar.penalty bug introduced in previous release. Added src/tests/test.earthmain.gcc.bat More man page tweaks 1.2-0 Added linpreds, allowed, and Use.beta.cache arguments Anova decomp is now more consistent Added a few GPL headers Reinstated the beta cache More man page tweaks 1.1-5 Fixed bug reported by Joe Ritzer: long predictor names got munged in plotmo Changed "class" to "response" throughout when used for the predicted responses in the input y. Man page tweaks based on user feedback. 1.1-4 Changed as.matrix to data.matrix in earth.default -- grep for FIXED Extended earth.Rd slightly 1.1-2 Added my web page to DESCRIPTION and to some man pages 1.1-1 Changed \r\n to \n to pacify CMD CHECK 1.1-0 Default minspan is now 1 (was 0) Fixed potential crash in PrintForwardStep if nTrace>1 Added a missing drop=FALSE to backward() Minor code, comment, and man page fixups 1.0-8 Fixed bug where plotmo failed under these circumstances: form <- Volume ~ .; a <- earth(form, data = trees); plotmo(a) 1.0-7 Minor change to summary.earth formatting. Man page fixes. 1.0-6 May 11 2007 Initial release