TODO:
o reduce/remove/rewrite for smaller code base
  * lr.c: combine update functions
  * spardat.c: reduce ivec/dyv/dym calls
  * ivec.c, amdyv.c: make simpler static classes
  * ...
o consider alternate sparse vector formats, maybe Judy trees
o consider using abdat, once they are finished and if I can get permission.
o remove AMFAST

CHANGES:
2006-05-18:
o RELEASE
o Increased line buffer size in amdym.c (used for reading csv files).
o Added extra checks and error messages to warn user if their file has
  lines longer than the buffer.
o Fixed comment in lr.c.
o Added verbosity check for predict's AUC output.
o Cosmetic Makefile change for release build.
o Improved error messages for case where the test point has
  fewer or more dimensions than the data.

2005-08-16:
o Removed wait_for_key() code.  my_error() will now simply exit.

2005-08-15:
o Spardats automatically size their width to the highest nonzero
  attributes seen.  A training spardat and a testing spardat will have
  different sizes when the highest attributes are unused in one or the
  other.  If the training set is wider than the testset, there is no
  problem when making predictions (never ask for higher coeffs in sum).
  When the testing set is wider, then we would try to deref non-existant
  model coeffs.  We now shrink the training spardat to the size of the
  model coeffs, and print a warning.  We did *not* do this for
  CSV files.
o Streamlined dyv_partial_sum when DEBUG is defined.  I don't have
  time to test the changes, so I don't know that it truly makes a
  performance difference.
o Added DEBUG define for debug and fastdebug builds.

2005-07-19:
o RELEASE
o Fixed wrong dates in CHANGELOG.
o Tried to fix zlib problems under cygwin, but gave up.  However,
  I will keep the file-pointer-ref-and-cast macros for pfiles.
o Removed zlib from mingw build type.
o Fixed UZLIB typo in Makefile.conf.
o Increased line-size limit in am_string:mk_string_from_line().

CHANGES:
2005-07-09:
o RELEASE

2005-07-08:
o Added zlib support everywhere.
o Fixed some small errors in the documentation, added short notes about zlib.

2005-07-07:
o Fixed bug in mk_lr_train() where return code from lr_train_iterate()
  was improperly ignored.


CHANGES:
2005-06-13:
o RELEASE
o Changed dense dataset load to reduce memory overhead.
o Changed lr_train stuff for dense datasets to reduce memory overhead.
  No dataset copies are made for training and predicting on dense data.
  k-fold cross-validation makes only one copy now.  Empirical memory
  usage during dense train is only slightly higher than dataset size
  times sizeof(double).  Dense k-fold is twice this.

2005-06-09:
o RELEASE
o Fixed a few build system bugs
o Changes to build system to accomodate Windows, cygwin and mingw

2005-05-24:
o RELEASE
o Added kfold executable
o Update docs
o Fixed cgeps description in docs
o Add valgrind target to testing makefile
o Set cgdeveps (+cgbinit, implicitly) as default cg termination method
o Fixed unitialized variable problem with cgbinit
o Removed several memory leaks (all known memleaks, using valgrind)
o Adjusted lots of verbosity-related stuff
o Added ROC-curve writing to predict program
o Thankfully, the algorithm times on ds1 are the same as those we have
  reported elsewhere -- at least we have not (yet) really screwed things up

2005-05-23:
o Big changes to build system

2005-05-22:
o RELEASE
o Initial release