Re: [Bug-gnubg] Is eval-4ply and 0-ply-rollout-truncated-4ply the same?
From:
Massimiliano Maini
Subject:
Re: [Bug-gnubg] Is eval-4ply and 0-ply-rollout-truncated-4ply the same?
Date:
Wed, 28 Nov 2007 13:46:58 +0100
address@hidden
wrote on 28/11/2007 11:07:57:
> On gammonu Michael Depreli asked,
>
> Is truncate at 4 the same as a 4-ply eval assuming you roll out at
0-ply ?
>
> I thought that the answer was an obvious yes, and Raccoon did some
> tests showing that the results where almost the same, but not quite.
>
> Jordan Lampe seems to think differently:
>
> >No.
> >
> >First of all, it makes a difference whether we are looking at
a
> checker play or a cube decision,
> >but let's suppose it's a cube decision just to make things easier.
> Let's further make things
> >easier by assuming there are only two die rolls (1 and 2) and
only
> two possible moves (A and B) at each step.
> >
> >A 2-ply evaluation looks at all of the following 16 positions:
1A1A,
> 1A1B, 1A2A,
> > 1A2B, 1B1A, 1B1B, 1B2A, 1B2B 2A1A, 2A1B, 2A2A, 2A2B, 2B1A, 2B1B,
2B2A, 2B2B
> >
> >A 0-ply, trucated at 2-rollout, on the roll sequence 1 followed
by
> 2 does this:
> >first it looks at 1A and 1B and decides which is better. Assume
1A
> looks better,
> >then it looks at 1A2A and 1A2B and decides which is better.
> >
> >So, in particular, the truncated rollout NEVER looks at 1B1A,
1B1B,
> 1B2A or 1B2B, no matter what dice it rolls.
>
> This seems horrible to me. Anybody know what is actually going on
or
> do I have to go look at the code? :)
Isn't this simply the fact that when doing a rollout
gnubg plays optimally against itself ? For a given roll, all the moves are
evaluated (at the specified ply), but only the best one is "played"
(i.e. kept in the rollout process). This is because the equity after a given roll is the
equity of the position you reach playing the move optimally. Hence whatever happens
in other branches does not affect that equity.
Still, a 0-ply rollout truncated to 4 is not the same
thing of a 4-ply eval. But if our neural net is good, the results shouldn't
be far apart in most cases, right ? I would expect the 4-ply eval to be
more accurate.