Author: José Carlos
Date: 08:30:19 11/10/03
Go up one level in this thread
On November 10, 2003 at 11:11:00, Daniel Clausen wrote: >On November 10, 2003 at 09:56:38, José Carlos wrote: > >[snip] > >> Now a position with queens on the board migh have 20 interesting moves. The >>engine will be able to search 10 plies in the same time. (I'm making up the >>numbers just to show the idea). >> So if there's a way to transpose to a pawn ending, and the win can be seen 15 >>plies after the pawn ending happens, a search from the root position won't be >>able to see it. >> You can perform extensions that say "if the tree is very thick in some place >>then extend depth". This is done for example when in check. If I'm in check I >>have very few moves to get out of check (usually) so I extend the depth. >> But you can't (or at least I don't know how to do it efficiently) search a >>move very deep before any other move, because if you run out of time searching >>it, and it proves to be bad, you have no idea what to play. > >I see the problem but still think it's an interesting thought. If we'd somehow >know that a certain subtree is "easier to calculate", we could extend the >subnode. I see two questions here: > >(1) >How to we know/guess that a certain subtree is easier to calculate than another >one? I can see two ways to do this. (a) With reduced material on the board, it's >typically easier to calculate (b) With some additional information from the HT >(like average branching factor of the subtree or something) > >(2) >While it's clear that in the given position it would be a good thing to spend >some more time in the material-reduced subtree after the queen/rook captures, >I'm not sure that's also the best in general. I could imagine it would be a good >thing though. > >I probably overlooked something obvious, but nevertheless the idea is on my >'ideas to try' list. :) > >Sargon I've also thought about it, but with no clear conclusion yet. Less nodes in a subtree usually means "easy to refute" subtree, and thus uninteresting. In a few cases, like when going to a pawn ending, it means a drawstical change in the shape of the tree. Note that reducing material doesn't necesarily make the tree thicker. changing some pawns and the knights might leave an open board for queens and rooks and increase the number of plausible moves. Closing the pawn structure makes the tree easier to search, but also less interesting due to less tactical chances. Simplifying to a pawn ending should be extended if if makes sense. If I'm a queen up and I give it for free it's obviously absurd to extend, though some may argue "extend always and prune later if necessary". I think (and I'm working around this idea) that there're posiotions that need search more than others. We need to figure out when an static eval is enough (KQK ending, I know this is won so no need to waste a single node here) and when there's a big need for searching (some pieces en-prise; king under attack). Then, every new iteration could increase depth differently for different type of positions. José C.
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.