[developers] Re: estimate and [incr tsdb()]

[Erik Velldal]

On 8/31/05, Stephan Oepen <oe at csli.stanford.edu> wrote:
> hi yi,
> 
> > A closer look at Itsdb and MaxEnt shows that there's a problem with
> > the "-variances" argument. Since it's not documented in MaxEnt manual, I
> > have no clue what should I do with it. Is it safe if I just remove it?
>

hello yi,
in general the file name specified for the -variances option should
contain either a single number that then sets the global variance for
the priors (this is the set up used in maxent.lisp), or one number per
line, individually specifying the variance for the prior used for each
single feature weight (see the note from rob on
http://www.isi.edu/~graehl/maxent/malouf/README.malouf.update).

in [incr tsdb()], setting the *maxent-variance* variable to NIL
instead of a number, simply means that estimation is done without
using a prior (i.e. the -variances option isn't passed to 'estimate'),
and this should work fine. using a prior on the weights can help
reduce overfitting, but you can estimate a model without it.

where exactly did you see messages indicating that it was the
-variances argument that caused problems? (i couldn't seem to find
anything related to that in the output and buffer files you sent).

cheers,
-erik