Author: Volker Böhm
Date: 09:07:42 07/02/04
Go up one level in this thread
On July 02, 2004 at 09:21:41, Fabien Letouzey wrote: >> Hi all, > >> This is my first post to this forum [I have been reading it for some time] > >> Some questions in arbitrary order: > >> 1. should PVS kick in after 1st move, or after PV found as Bruce suggests on his >> page? [I *think* tree is smaller Bruce's way] > >It is the first time I have a look at Bruce's example. > >In classical PVS/NegaScout, use the scout search after the first move. >The assumption is that the first move is likely to be the best one. >This applies even if the score is below alpha. > >Bruce makes a slightly different assumption that depends on the >window. IMO it is an uncommon implementation but it might be better >depending on aspiration windows. What does Bruce suggest? Btw. who is Bruce, Bruce Moreland? > >I suggest you start with the classical way. > >> 2. if using fail soft, which is right: pvs(-best-1, best) or pvs(-alpha-1, >> alpha) ? > >Note that you forgot the minus signs for both second arguments, the >windows should be "minimal" (beta = alpha+1). Anyway, only testing >against alpha is correct. Otherwise, you might end up using an even >larger window than plain alpha-beta during the re-search. > >> 3. depending on 2., how do you test for failure, and what bounds to use for >> research? best or alpha > >After a minimal-window search like "v = search(x,x+1);", a fail low >is characterised by "v <= x" and a fail high by "v > x". Just be >consistent between the search call and the subsequent test. > >Usually after a standard NegaMax "v = -search(-alpha-1,-alpha);", one >tests for a fail high with "v > alpha" (the two minus signs cancel out). > >> 4. Fruit uses 'mate distance prunning' - I read in the archive [now down], >> where some suggest using shallower mate score hash entries, and others suggest >> boosting the 'draft' of mate scores [which is similar, I suppose] >> Are these all different versions of the same thing? - any help? > >The last two sound like the same. A trap to avoid: an "exact" score >of "mate in 4" at depth 5 might become "mate in 3" at depth 7. So >when you "boost" the draft, make the score a bound. > >The "mate-distance pruning" I use is common, but I don't know a >standard name for it. It is a completely different idea, that is >actually independent of the transposition table. If at a given node >you already have a "mate in 3" (= 5 plies) then only a "mate in 2" (= >3 plies) or shorter could improve the current value. Therefore there >is no need to look more than 3 plies ahead in any subtree. You might >need to adjust depending on how you detect mates. Also most engines >count mate distances "to the root", so implementation requires care. > >> Thanks for *ALL* your help: past, present, and future; > >Just help back when your turn comes. > >> David > >Fabien.
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.