Next: Glossary of Frequently Used
Up: METAL The METAL Machine
Previous: Structure and Organization of
  Contents
Subsections
If something goes wrong, always carefully look into the log file.
If there is no hint what went wrong, repeat the experiment with
the added options -d and -lad. This will
create a huge logfile, since all the output of all
learning algorithms will be included, but usually contains
the crucial information about what went wrong.
Some things to check for: are all programs that are needed from
the interface scripts in the binary search path? The helper
programs too? Is the directory used to store temporary files
on a device that has enough free space to hold all the temporary files?
Are there leftover temporary files from earlier runs that clobber
up space?
On some systems - mainly Windows - the CPU limitation mechanism does not
work. Unfortunately there is no solution for this as of now.
Apart from the cause given in the previous problem, it is also possible
that the learning algorithm or some other algorithm that gets called
(indirectly via interface scripts) from run_exp
is waiting
for input from the standard input stream. In that case, the algorithm is
halted, does not consume CPU time and thus, never stops.
One reason for this behaviour could be that a licensed learning
algorithm is requesting a license code.
If the database is very small and there are many different class labels,
the shuffle
program will not
be able to do stratification correctly without leaving some of
the files empty. In that case, simply turn off stratification
(option -start 0
).
Next: Glossary of Frequently Used
Up: METAL The METAL Machine
Previous: Structure and Organization of
  Contents
2002-10-17