bug-gnubg
[Top][All Lists]

## RE: [Bug-gnubg] RE:Why is odd ply equity always lower?

 From: Ian Shaw Subject: RE: [Bug-gnubg] RE:Why is odd ply equity always lower? Date: Thu, 12 Jun 2003 15:42:10 +0100

```Joseph Heled wrote:
> Ian Shaw wrote:
> >
> > Since a 1-ply eval is just 21 0-ply evals, it appears that GnuBg is
> > consistently overvaluing the equity for the side on roll.
> >
>
> The net *always* evaluates for the side on roll. My guess is
> it has to
> do with the selection of positions. I think it is unlikely to be true
> for a random position. Can anyone else confirm this?
>

I think I haven't been precise enough. I understand that the net always
evaluates for the side on roll.

As I understand it, on a 1-ply evaluation of a position, the net is fed all 21
possible return rolls for the opponent, and the equities of the best moves are
averaged. The equity for the original player is then one minus this average. If
the net tends to overvalue the equity of the side on roll, then the 1-ply
equity will tend to be lower than the 0-ply equity. We would then expect this
trend to continue for even and odd plies.

Using the 0.14 net, I have just evaluated 40 plays (money play) of the opening
roll on 0-ply through 4-ply, and have observed this trend consistently.
Typically, the odd-even discrepancy is about 0.05, which is fairly significant.
It begs the question, which is closest to the correct equity, and would we get
different results for rollouts at different plies? At present, the consensus at
GammOnLine is to avoid gnubg odd-ply evaluations and rollouts, and I can't see
this changing while the odd-even discrepancy is so marked.

I can't recall any evidence that the even plies are better - I think it's
mainly a case of 0 and 2 ply ganging up on 1 ply. Most people don't bother with
3 and 4 ply comparisons.

Someone (Nis?) was looking at combining results to get 1/2 ply evaluations. Is
that relevant here. I guess it's a bad idea, seeing that odd & even plies are
not on the same wavelength.

I have the opening roll results as a text file, if anyone wants them. The
interesting results were the usual suspects:

Opening 32 - Split 24/21 13/11 at all plies except at 0-ply (build)
Opening 41 - Split 24/23 13/9. 13/9 6/5 loses favour at higher plies
Opening 43 - Long split at 0-ply, short split at 1-ply, build at 2 to 4 ply
Opening 52 - Split 24/22 13/8 except at 2-ply (build)
Opening 64 - Build 24/18 13/9 at even plies, point at odd

--Ian

```