repetitive training

[Boris 'pi' Piwinger]

Greg Louis wrote:

> I think there is still some doubt possible regarding the benefits of
> repetitive training, and it mightn't be a bad idea to mention
>     http://www.bgl.nu/bogofilter/training.html

This test does not make use of security intervals, which
would be useful. Also you don't get to closeoff. But this
tests certainly comes closest to training to exhaustion.

>     http://www.bgl.nu/bogofilter/training2.html and

As far as I understand this does not try repetitive training
on error runs. You just train on error with new messages,
i.e., not using the old messages again.

>     http://www.bgl.nu/bogofilter/reptrain.html

As I said earlier, this test says nothing about training to
exhaustion, it violates basic assumptions. First you do some
initial full training (else bogotune would intentionally
refuse to work), then you change parameters. It is unclear
what these parameter changes are, but this produces complete
unpredictable results. If bogotune suggests to get many fp's
than this seems to be a bad suggestion. It would be very
important to see a) what the parameter changes are and b) to
see before and after parameter changes the number of errors.

pi, small p