[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Bug-gnubg] Suggestion: equivalent # of games played in roll outs th
Re: [Bug-gnubg] Suggestion: equivalent # of games played in roll outs thx to var. reduction
Tue, 14 Oct 2003 19:12:44 +0200
On Tue 14 Oct 2003 (13:23 +0000), Joern Thyssen wrote:
> On Tue, Oct 14, 2003 at 02:49:15PM +0200, olivier croisille wrote
> > I'm aware that the issue was raised in the past various times, I'd just
> > like to know the current status, if any :
> > - is it on the agenda at all?
> yes, it's noted on the TODO list.
> > - is it on the agenda but with a very low priority level? (which would
> > perfectly make sense, regarding a more or less cosmetic feature)
> noone is working on it
> > - is it hard to compute
> To compute it you need the ratio between the calculated standard error
> without variance reduction and the calculated standard error with
> variance reduction. The equivalent #games varies with the square of this
> ratio, e.g., a 36 game rollout where the SE without and with VR is 0.01
> and 0.005, respectively, the #equiv games is 4*36=144.
> > implement?
> yes :-)
> I'm guessing that we need to save the non-VR equity and SE (internally
> and externally) in order to be able to resume rollouts. This is the
> critical point, as it requires changes to many places in gnubg.
Yes. Even if we decided to print 'many' instead of maintaining the
history for resumed rollouts, we'd need to change RolloutGeneral (not
too hard), the moverecord analysis struct to record the equivalent
games (and the non-VR results if we wanted to be able to resume), then
re-format every display of the results, pdf, html, text, to show this
> > Incidentally, on the 'promotion' side (if ever needed) I think this would
> > show even better the top quality of roll outs enabled by Gnubg
> Personally I don't find the #equiv games that interesting. The SEs and
> JSDs are much more interesting -- it doesn't help you much to know that
> your rollouts is equivalent to 10,000 games if the two moves you're
> rolling out are only separated by one tenth of a SE.
> All rollouts are normally performed with VR, so what's the point of
> knowing that the #equiv games for a non-VR rollout, since nobody would
> do such a rollout (unless they only have S3.2 and enable chequer play
> according to score).
> Other opinions on this subject?
I think we should simply document why we don't do this. Being told
that gnubg has done BIGNUM rollouts may cause people to believe that
two moves separated by .6 j.s.d.s are well ordered, simply because the
number is big.
Jim Segrave address@hidden