Theo Van Dinter writes: > On Mon, Jan 01, 2007 at 02:43:25PM -0500, Daryl C. W. O'Shea wrote: > > This is the biggest reason why I'm not yet running 3.2 on my MXes... > > there's nothing to stop rules with inappropriate scores, or rules that > > overlap (all with a score of 1, probably) from being auto-promoted and > > published. At least with the manual updates someone has to consciously > > add those rules and their scores to the update. > > This is my main issue with the automatic updates btw. I currently don't like > the idea of the nightly/weekly results enabling/disabling rules and changing > scores (though it doesn't do that yet iirc) on production-use updates.
My current thinking is that we can lock the scores of the "main" ruleset, so they cannot be changed, and use the perceptron to generate scores for the new rules. If update rules overlap with each other, or with the main ruleset's scores, this should take care of that (by either setting some to 0 or giving them all low scores). It might also be wise to run the perceptron with a lower threshold than usual, so that FPs are kept artificially low. Also, some manual score-setting should still be possible (for cases when we know that a rule needs to score higher or lower than the perceptron would guess). The current situation -- all-manual score setting -- is suboptimal, btw. I think that quite a few of our rules are overlapping -- so even though we *think* we're setting low scores, due to 3 rules all firing on the same mails, the end effect is that certain message traits will score highly. The perceptron will deal with that problem better than we do. --j.
