Re: [Bug-gnubg] Odd-Even evals - checkerplay

[nis . anders . jorgensen]

Sending this again , since it only went to Henrik the first time ...


Henrik Bukkjaer wrote:

>Hi all fellow BG freaks!
>
>Just back from vacation, I'm 400+ mails behind in this forum...
>
>Anyway, I have also noticed the odd-even ply evaluation variations that
>have been discussed, but unfortunately I have not had the time to
>store/analyze the findings. I will start to collect them from now on.
>
>My question:
>After having discussed the problem, do you have any explanation why the
>systematic variance occurs?

I posted my theory on the group some time ago :

http://mail.gnu.org/archive/html/bug-gnubg/2003-06/msg00207.html

This does NOT seem to fully explain the position you are posting,
however - since that one has one of the 3-ply equities being higher than
2-ply. My guess is that what we are seeing is a combination of a
"random" disagreement between odd and even, and the systematic bias of
odd plies being evaluated lower than even ones which I described.

>It has been suggested here to use "half-plies" to correct the problem (to
>give sort of an interpolation).
>To me this seems wrong - even if it would give more accurate evaluations.
>It is like treating the symptoms as opposed to find the root cause and
>completely cure the gnu.

Well, if my theory is correct, this is an "inherent" problem in all
bots. Thus a fix would have to radically rethink the way deeper-ply
evaluations are done. The "half ply" approach actually "cures" the
systematic bias to some degree - by giving an alternative evaluation,
which evens it out.
At the same time, it serves as a "dampener" on any disagreement between
odd and even - though there is of course no guarantee that the truth is
in the middle.

>Also it has been suggested that the odd-even problem is not important for
>checkerplay but only for cube decisions.
>This is not correct.
>The odd-even problem only applies to some kind of positions. For some
>checkerplay decisions you must decide between two checkerplays leading to
>very different gameplans. The odd-even problem could apply to only one of
>these, resulting in gnu choosing play A on 0-, 2- and 4-ply, but choosing
>play B on 1- and 3-ply evaluation.

Correct.

>Pos: zG2DARjYm4URIA
>Match: UQkpAgAAIAAA
>(I was using 0.13b on the Ortega & Kleinman 2001 ME table)
>
I would suggest posting board AND match ID/pos ID - since I haven't
really gotten used to reading positions from postision ID's, and
sometimes read email in places with no gnubg installed.

>The position shows a very interesting checkerplay decision - go for the
>prime (13/7 9/7) or go for the stronger board and fewer blots (9/3 5/3).
>(The roll of 6-2 leads to the same checkerplays as the actual roll of 2-2).
>
>Gnu chooses the prime move on all plies.

Just FYI, 0.14 with Zadeh MET chooses 9/3 5/3 on 0-ply

-- 
Nis Jorgensen
Your man in Amsterdam