[Top][All Lists]

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

Re: [Bug-gnubg] patches for changing evaluation in mid-rollout

From: Jim Segrave
Subject: Re: [Bug-gnubg] patches for changing evaluation in mid-rollout
Date: Wed, 30 Oct 2002 23:39:33 +0100
User-agent: Mutt/

On Wed 30 Oct 2002 (01:22 -0800), Ned Cross wrote:
> Hi,
>     The user interface looks great at first glance.
> I just set up a 2-ply, 2-step rollout of yesterday's GOL cube decision, and
> the equities for "player xyz owns 2-cube" appear to be doubled.
> After 36 trials, GNU is giving a cubeless equity of ~ +2.500 for double/take
> (61% wins, 33% gammon wins, 0.3% bg wins, 12% gammon losses, 4% bg losses)
> when my 0-ply rollout gave +1.290, and 2-ply eval gave +1.499 (similar
> win/loss/gammon breakdown).

I saw your followup where the same rollout was settling to a
reasonable figure. I don't think my changes will have broken anything,
but I hope people will try it for a while to see what they think.

As to why this might have started off so odd...

Are you doing a 1296 trial rollout? If so, then the first 36 trials
will be for a roll of 1-1, which for Kit would have been a nice roll,
he could close his 5 point and put a 2nd chequer on the bar, leaving
the readers with two to bring in against a 4 point board. So the
equities on 1-1 would be pretty high I would think. By the time you
get to 770 rolls, you're looking at 4-5 or similar and some of the not
so good rolls with good followups from the reader's side will have
brought things back down.

If you aren't doing a 1296 trial, then I don't know. 
The behaviour with rotate rolls and 1296 games is the point of some of
the discussions about randomising the order of doing the trials in a
rollout. If we randomise, then stopping a long rollout (or simply
doing what you did and noting the numbers while it progressed) are
less likely to produce such huge swings in apparent equity.

Jim Segrave           address@hidden

reply via email to

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