Re: [Bug-gnubg] Oddity/Bug in Rollouts?

[Christian Anthon]

On Mon, Apr 14, 2008 at 2:19 PM, Massimiliano Maini
<address@hidden> wrote:

> - how can a position be too good for a player that can not score a gammon ?
> Well, he could get backgammons, but it doesn't seem to be the case above.

The most likely cause is that O makes some bad takes on the next roll.
That would send your cubeful expectations above 1.0, normally reserved
for too good. This turn the possition appears to be a simple double
take. Next roll it could easily be a double pass. It is mostly a
matter of matching numbers and logic.

>
> - how can a position be too good for player A and be a take for player B ?
> I thought that, by definition, a position is too good when my (normalized)
> equity is grater then +1.000. In this case, if doubled my opponent should
> drop, losing only -1.000. Too good/take does not exist.
>

It does when you are playing a match and gnubg uses the same naming
conversions for money and match. The only difference is that in the
present position you are too good because your opponent is a fool and
not because of gammons.

> Maybe what's happening is this: rolling out the nodouble part, X doubles
> later on and O takes. This can explain how X nodouble equity is +1.015 >
> +1.000.
>

Exactly.

>
> But the whole odd/even thing looks evil to me. For example, take this
> position (same as previous but with closed board) with black/X on roll:
>
>  GNU Backgammon  Position ID: BwAA2LZtAUEAAA
>                  Match ID   : UQngARAAIAAA

First of all you are better of comparing static evals with static
evals when you are comparing nets. It might not matter in this case,
but in general you are more sure of what you are doing. Second of all
the 0.720 is reproduced as 0.280 in 1ply evals.

> Now put white/O on roll: it shouldn't matter since white is on the bar
> against a closed board, but ...

This assumes a dynamic view of the position which the neuralnet
doesn't have. It doesn't see that the position is independent of the
36 dice rolls. Even if it did what should it use this information for?
It would make no sense prefer the 1ply over 0ply eval for a particular
case. All in all the 0ply/1ply phenomenon could be as large in this
position as in the original example. Only the quality of the nets in
the individual positions really determine the size of the effect.

>
> P.S.
> I start gnubg, open un the following sgf (position):
>
>         http://www.gnubg.org/media/odd-pos-closed1.sgf
>

Reporting bugs as a ps. to a message doesn't exactly make it easy to
keep track of the bugs that cannot be fixed at a glance. Please put it
up on savannah where you may attach the file as well.

Christian.