MLC++: A Machine Learning Library in C++                     11/21/97
Cliff Brunk

                How to compile the MLC++ Sources under g++
                -------------------------------------------

These notes are designed so MLC++ can be compiled on SUN using GNU
GCC 2.7.2.3 or on Silicon Graphics using GCC 2.7.2.2.


** Please note that if you only want to work with MLC++, you can get
   the MLC++ utilities off our home page.  There are precompiled versions
   for SGI, SUN, and NT, these are much easier to install and run.  

If you have any questions, send mail to mlc@postofc.corp.sgi.com.



Compiling MLC++ under GNU
-------------------------
*. Install the GNU gcc compiler.
   A precompiled executable version of GCC 2.7.2.3 for Solaris 2.5.1
   can be found at:

   ftp://sunsite.unc.edu/pub/solaris/sparc/GNUgcc.2.7.2.3.SPARC.Solaris.2.5.1.pkg.tgz

*. Get the MLCX.Y-src.tar.gz file from the MLC++ home page, where
   X.Y is the version number.

*. Create a top level MLC++ directory and make that the current working
   directory.  (e.g. cd; mkdir mlc; cd mlc)

*. gzcat MLCX.Y-src.tar.gz | tar xf -

*. Set MLCDIR to the top level MLC++ directory (e.g. setenv MLCDIR ~/mlc)

*. Setup the MLC++ environment variables for the SUN (source setup.SUN.GNU)
   It is recommended that you put this in your .login

*. Optional: Install dot/dotty in subdirectory graphviz
   Go to http://www.research.att.com/sw/tools/reuse/
   click binary software license agreement and mark graphviz.

*. Set your PATH to access: ar, ld, strip, as, gcc, g++, dot, dotty and
   all the external inducers that you will be using.
   
   The external inducers supplied with MLC++ (T2, cn2, ibl, oc1, pebls)
   can be built using the command:

   	bin/buildexternal T2 cn2 ibl oc1 pebls

   The corresponding executables are stored in

        ${MLCDIR}/external/${MLC_EXTERNAL_TYPE}

   The following commands will modify your PATH allowing access to
   dot, dotty and lefty in the graphviz directory and the external
   inducer products stored in ${MLCDIR}/external/${MLC_EXTERNAL_TYPE}:

        setenv PATH "${MLCDIR}/graphviz/bin:${PATH}"
        setenv PATH "${MLCDIR}/external/${MLC_EXTERNAL_TYPE}:${PATH}:."

*. Some testers use c4.5, c5.0, cart and dotty.
   These testers will fail if you do not have the products installed.
   If this is the case, simply do "touch t_XXXX" where t_XXXX is
   the name of the tester that failed (no suffix).  The make
   will then ignore that tester.

   Note that these will be removed by a "make clean" or "make scratch"

   The verifiers for utilities (mlc/util dir) will also warn you about
   diff failures. If these are related to C4.5 and the other tools,
   this is expected.

   Note: MLC++ provides an interface to cart, however it requires that
   you rename the ascii conversion program from ascii to cart.ascii.

*. Build MLC++

   make - in the top level MLC++ directory builds and of the MLC++ libraries,
          utilities and testers

   make notests - builds the MLC++ libraries and utilities without testing

   make tests - builds the tests after the libraries have been built

   If something fails, you can type check the appropriate
   execution log in the tester name with a ".out" suffix.
   For example, the first file to fail if you don't have
   C4.5 is $MLCDIR/src/MTree/tests/t_C45Inducer.  The t_C45Inducer.out
   file will contain the error "c4.5: not found."

   To continue after you fix a problem, type "make" in $MLCDIR/src
   or in any subdirectory.  Use "make -r" when doing a make in the
   tests directories.  The  -r is required to prevent implicit rules
   from doing the wrong thing.

   The default is to compile all the library and run all the testers.
   This is a long process

   If you have to clean something, do it in a subdirectory.
   No need to do it at the src/ level.  If the compilation
   failed in src/MGraph/tests, make clean in MGraph.

*. If you develop code, we highly recommend working in DEBUGLEVEL 1
   until the code works well.




Known MLC++ problems on SUN
---------------------------
All of the known problems are related to testers and we are currently
working on solving them.  Most involve differences in real number precision.
The following is a list of the known problems:

src/MTree/tests/t_RDGCat
	binary files differ (probably a SUN binary file issue)

src/MTree/tests/t_treeviz
	precision difference in t_treeviz.exp14

src/MInd/tests/t_NaiveBayesCat
	precision difference t_NaiveBayesCat.exp3

src/MInd/tests/t_TableInducerMTrans/tests/t_PeblsInducer.tmp
	problem matching instances because of Real ==

src/MTrans/tests/t_CN2Inducer
	expected:	ERROR : 0.2
	observed:	ERROR : 0.21

src/MTrans/tests/t_MultiSplitCat2
	doesn't generate output files .out1, out2, out3

src/MTrans/tests/t_SGIDTInducer
	t_SGIDTInducer.exp7 generates a different set of models under GNU

src/MFSS/tests/t_TableCasInd
	expected:	ERROR : 0.633191489362
	observed:	ERROR : 0.632553191489





   Cliff Brunk (brunk@sgi.com)


====================