[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] Problems with stopping rollouts when STDs are small enou

From: Massimiliano Maini
Subject: Re: [Bug-gnubg] Problems with stopping rollouts when STDs are small enough
Date: Thu, 5 Feb 2009 14:09:26 +0100

address@hidden wrote on 04/02/2009 00:14:32:

> One of the available criteria to end a rollout is when standard deviations
> are small enough. For instance, it could be :
> Stop rollout when STD's are small enough
> After 1296 games, rollouts will stop if the STDs are small enough.
> Rollouts can stop when the ratio |STD/value| is less than 0.0040 for
> every value (win/gammon/backgammon/...equity)
> I think there are two problems with the way this is curently implemented :
> - the "ratio ... for every value" works poorly when one of the values,
> usually backgammons for one of the players, is small but not zero. All
> other STD/value ratios are roughly identical but that of this irrelevant
> (equity-wise) term can be 10 or 20 or 30 times larger and cause the
> rollout to take much more time.
> I'd suggest we check only the equity. If it is accurate, the single
> win/gammon/backgammon numbers will be as well if they are any significant.

Makes much sense to me. I was even tinking the stopping criteria was already
working in the manner you describe.
> - if one doesn't care for the individual kind of outcomes, the STD/value
> ratio doesn't seem pertinent. Do we really want to know equities around
> 0.2 with an accuracy five times higher than those around 1 for instance ?
> The simple STD on a 1-cube, whatever the equity, seems a better criterion.

I suppose that if one is not interested in estimatnig individual equities
then his/her interest is in estimating the relative ranking of the alternatives.
In this case, the jsd-based criteria seems better.

Anyway, I've never digged too much into these options, I just wait for the
unstopped rollout to finish :)


reply via email to

[Prev in Thread] Current Thread [Next in Thread]