Re: [Bug-gnubg] I'm a bit confused

[Jim Segrave]

On Tue 29 Jul 2003 (21:13 +0000), Joern Thyssen wrote:
> On Tue, Jul 29, 2003 at 04:19:01PM -0300, Albert Silver wrote
> > 
> > I'm trying to understand why move 3 is a -0.147 blunder when its wins
> > are only 0.1% less, it scores more gammons, and loses less gammons and
> > backgammons. 
> 
> You can't rely on the cubeless numbers when doing a cubeful rollout.
> 
> If the game reaches a "double, pass" then the rollout is truncated with
> a cubeful equity of +1 (or -1) and gnubg uses the cubeless numbers from
> the cube decision (which is 0-ply in your case). 
> 
> Worst case scenario: you're on roll in a position where you'll be
> doubled out no matter what you roll. If you do a 0-ply cubeful rollout
> then the cubeful equity is -1 and the cubeless gwc is identical to a 
> 1-ply evaluation even though you've selected 1,679,616 trials.
> 
> For this very reason I've been on the verge of removing the output
> wins/gammons/backgammons from all evaluations and rollouts! Your
> question is not the first, and certainly not the last! 
> 
> Try doing a full cubeless rollout. I imagine that you'll see totally
> different numbers. Your numbers may also be an indication of 0-ply not
> being good enough for the cube decision, so you can try a rollout with
> 0-ply chequerplay and 2-ply cube play.
> 
> You may see similar problems for evaluations as well, but as evaluations
> are never truncated then the problem is usually not so large.

Now I'm confused - I would think that the wins/gammons/backgammons
figures are accurate, but probably not complete enough. In the
situation you describe where you'll be doubled out, your wins should
be 0.0, lose gammon and lose backgammon will be either 0.0 or 1.0
depending on where your pieces are.

I would find the cubeless rollout figures just as strange, since they
would show results which would never happen when the cube is live -
even in a situation where you're going to be doubled out, if it's
played out you could expect to win 5% to perhaps as much as 30% of the
games, depending on the current match score and where the cube is at
the time the doubling out takes place.

If rollouts of moves kept statistics, then it would be useful to look
and see where the wins and losses actually occur, since that's the
only reason I can see for Albert's puzzling results. Certainly many
users of gnubg, snowie, jellyfish et. al. routinely look at the
win/gammon/backgammon figures when trying to understand why the bots
prefer one move over another.

-- 
Jim Segrave           address@hidden