Author: Dieter Buerssner
Date: 16:50:59 08/07/03
Go up one level in this thread
On August 07, 2003 at 19:36:45, Omid David Tabibi wrote: >I think due to the search instability, the best thing is to use the full window >upon re-search, i.e., in the example above: > >if (value > alpha && value < beta) > best = -PrincipalVariation(pos, depth-1, -beta, -alpha); > /* ^^^^^^ */ Yes and no. I do it, as you suggested. My comments were to a plain alpha-beta search (no pruning, extensions, qsearch). The "PV-updating-problem" exists already under those conditions, however. But it is discussable, which is better (using -alpha or -value+1) as bound in the research. The first seems more logical at first sight (taking into consideration the search instabilities, you mentioned). I have tried both, and did not see much difference. I debugged very large search trees. There are some subtle points (for example, do you trust hash bounds and adjust bounds in subnodes, and more). I am not sure anymore, and tend to think, it is not worse to trust the first search. After all it said, the score is >= value. Why should we trust the research more (that, with your method could now say, hmmm, it is really smaller value, when I look again)? I think, this is no easy to answer question. I have the feeling, that commercial engines trust their older searches more, than amateur engines (including mine). >Another thing I find interesting is the two different implementations of PVS: > >I) Use full window for the first child, and minimal window (alpha, alpha + 1) in >all other childs (like the pseudo-code above. This is the PVS version used in >Crafty). > >II) Use full window for all childs until you find a PV, then use minimal window >for all other childs (see Bruce Moreland's pseudo-code). And even more variations are possible. One can try to guess, whether we await a fail high in the next node or a fail low (most nodes will not be exact). And dependent on the guess switch between I and II. I tried a bit, and nothing spectacularily found (not surprising). Regards, Dieter
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.