Author: Vasik Rajlich
Date: 05:43:40 09/25/04
Go up one level in this thread
On September 25, 2004 at 06:35:57, martin fierz wrote: >well, as promised here are questions 4-7 of a zillion. perhaps i should also add >that the train ride from zurich to brissago takes about 3 hours and that i have >a centrino laptop, so i need some input :-) > >nullmove questions! > >1) do you save "don't nullmove" in your hashtable? if yes, how much does it help >you? isn't this a tinytiny improvement, since if you do nullmove, and it fails, >you have to research a much larger tree? >you would be saving "don't nullmove" if your nullmove fails, right? > >2) related: do you save matethreat in your hashtable? is this not another >tinytiny improvement? > I was doing something similar when my engine was pure MTD (f) - I was storing the results of the null move search in the hash table. I don't do this in my PVS engine. Intuitively, in PVS, you should rarely revisit a node where the hash table entry won't cause an immediate cutoff, yet where you would have enough information to prevent the null move search. Keep in mind also that if you do try a null move which is fated to fail low, the wasted search should involve just one node - you should get a hash table hit on the first node of the null-move search. >3) some people advocate not to nullmove if the side to move has few (mostly 0, >perhaps also 1?) pieces. i realize that zugzwang is a problem for nullmove, but >is it a real problem as far as games go? isn't it possible that in pawn endgames >doing a nullmove would lead to a few embarassing losses, but help overall? > It could be worth testing, but intuitively I would guess that any null-move searching in K&P endings would be counterproductive. One of the things that makes null move work in positional positions is the assumption that having a free move is very good positionally - good enough to offset any horizon effects from the reduced-depth search. As soon as this assumption breaks, null move will wreck the positional level. Vas >4) i never thought much about verified 0-move pruning. but a post i read last >week made me think: i always thought i could do the following: > >if(pieces > 0) > if nullmove-fails-high > return value; >if(pieces == 0) // do a verification search > if nullmove-fails-high > verify result with a normal search with reduced depth. > if verification-search-fails-high > return value; > >now there was this post saying that this won't work because you have to disable >the nullmove in your entire verification search. i don't understand why that >would be? can somebody explain? > >cheers > martin
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.