MLC++ Updates                            10/6/96
                       _______________

SGI MLC++ utilities 2.0 are RESEARCH DOMAIN only.  This means they
can be used for research/educational purposes.  See
   http://www.sgi.com/tech/mlc/terms.html
for details.

Appended is a list of new features in version 2.0

--

   Ronny Kohavi (ronnyk@sgi.com)


              Main Differences Between MLC++ 1.3.2 and 2.0
              ____________________________________________

 - The distribution is compiled in FAST mode, which is about
   30% faster than non-fast mode (non-fast mode was the mode
   previously distributed).

 - The utilities distribution is given using dynamically shared
   objects that save space.  The compressed tar file is now
   about half the size of the last version.
       
 - Persistent categorizers are now supported.  Persistent
   decision trees and Naive-Bayes are implemented.  This allows
   a categorizer to be saved and later read in.  Assimilation
   code allows instances with more attributes than required
   by the categorizer to be assimilated and categorized.

 - Decision trees were improved as follows:

    - Decision trees now provide pruning in a way similar to C4.5.
      Branch replacement is not being done and the fudge factors
      (C4.5 adds 0.1 in certain places are not in the code).
      The MC4 inducer defaults to a setting very similar
      to C4.5's setting.

    - A new option, adjust thresholds, allow splits in decision
      trees to be adjusted to actual data elements as in C4.5.
      The option is implemented much more efficiently than in C4.5.

    - Gain ratio is supported as a splitting criterion.  This is
      implemented exactly as the C4.5 version (with all the hacks),
      so that except for unknown handling and tie breakers, the
      unpruned trees are the same.

    - More statistics are provided about the number of attributes
      and depth of tree.

    - Improved output for MineSet(TM) Tree Visualizer.

 - Naive-Bayes changes:

    -  Naive-Bayes uses a value of 0 as a default for
       NO_MATCHES_FACTOR, which is the value used when there are
       no records matching a given attribute value and label value.
       This was made to make the probability distribution consistent
       and (surprisingly?) results sometimes improve.  The previous
       default was 0.5 over the number of instances.

    -  Naive-Bayes now supports Laplace corrections.

    -  Naive-Bayes now outputs MineSet(TM) Evidence Visualizer format files.

 - The biasVar utility has been added for the bias-variance
   decomposition based on Kohavi & Wolpert ICML-96 paper.

 - NBTree described in Kohavi, KDD-96 can be setup
         with the following parameters:
      setenv INDUCER catdt
      setenv CATDT_LEAF_INDUCER naive
      setenv LBOUND_MIN_SPLIT 30
      setenv CATDT_CV_FOLDS 5
      setenv CATDT_CV_TIMES 1
      setenv CATDT_LEAF_NB_NO_MATCHES_FACTOR = 0.5
      setenv IMPROVE_RATIO 0.05


 - A ``dribble'' support was added to show progress on big files.
   Dribble is done for decision trees and discretization.

 - Unlabelled instance lists are partially supported.  The syntax
   is to say ``nolabel'' in the names file.

 - Options in OODG allow you to build an oblivious decision
   tree to determine the cumulative purity of chosen attributes.

 - Control-c and kill signals are caught and handled. 
   A cleanup of temporary files is done.

 - Governors remove attributes with too many values and avoid
   runs with too many label values.  These limits are artificial
   and can be increased by changing the appropriate options
   in the message.  With this change, dynamic projections
   of instance lists are allowed during reading (mainly
   an efficiency issue).