Author: Álvaro Begué
Date: 11:43:53 06/13/05
Go up one level in this thread
On June 13, 2005 at 10:11:11, Stan Arts wrote:
>On June 13, 2005 at 09:37:46, Claude Le Page wrote:
>
>>Some years ago , many considered that evaluation of a move has not the same
>>value if the depth is even than same eval at odd depth
>>Is it always true , or only to small depth , and what is the reason of this
>>difference?
>
>
>Hi,
>
>Those scoreswings of even/odd depths often come from the horizon of quiet
>("positional") search mostly. If you search to 8 ply, there will be captures
>and some special moves beyond that, but mostly not positional moves. So for
>instance, in the opening searching developing pieces end there. At 8 ply, the
>opponent got to move last, (so maybe eval returns a small advantage for him, he
>developed an extra piece) If you search to 9 ply, you got the last move in the
>tree, and you could now develop a piece. (so eval might be even again.)
>
>But it's hard to say where an evaluationvalue comes from when a program has
>things like Q-search, extentions, reductions, and so on.
>
>When there are some tactics, scores often stay the same for several iterations,
>(because Q-search or extentions picked something up, and regular search has to
>catch up.)
>I can see these even/odd scoreswings especially in the opening, where the
>scoreswing is often the value of a piece being developed. But when the game is
>well underway it's hard to notice anymore (because of tactics) , but it does
>also seem to happen a lot in quiet/blocked positions, because of the same
>reason as in the opening.
>
>Greetings
>Stan
I think that is a sensible explanation. Another situation where some programs
find this effect is in promotion races (at least my program had this problem
until I fixed it), where the value of a passed pawn increases with each rank it
gets to.
This even/odd scoreswing is particularly bad if one of the moves you are
coinsidering is extended for some reason (like being a check). You can find
yourself in the situation where you are going to think that giving the check is
your best move at even depths and that there is a better move at odd depths,
just because the search is stopping at boards with different parity in the two
branches. One solution is giving a few points to the side to move (half of the
swing would be great). I only implemented this solution for the pawn races (I
actually consider the passed pawn to be in rank 6.5, for instance, and I give
more points smoothly as you approach the 8th rank).
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.