Author: Robert Hyatt
Date: 05:55:39 02/09/99
Go up one level in this thread
On February 09, 1999 at 07:51:43, Bas Hamstra wrote: >Can anyone help, please? I really could use some good advice about extensions. > >Since I added extensions to my program it solves more WAC positions and much >faster. So I hoped my FICS rating would climb quit a bit. But alas, the opposite >happened. I know, I know, suite solving times don't say anything about qualitity >of play in real games. Yet you can't afford to miss too much basic tactics, and >that's what I try to use WAC for. > >I think the reason why the rating is so low now (1920 FICS) is that the >extensions take too much "horsepower". I think. Not that de search exactly >chokes, but sometimes, for example with open kings on the board, the engine >thinks a couple of plies less deep than normal. > >What is the optimal way to do extensions *right*? So it still quickly sees the >"standard mates" but without too much costs? > >I am thinking of doing the following: > >1. Don't limit one-reply-to-check at all. Just add +2 to Depth. > that will cost you dearly if you aren't careful. IE at the previous ply, where your opponent checked you, _he_ could have gotten an extension for recapture or whatever. If you extend more than one ply every ply, or more than two plies every two plies, you can have a search that blows up. Not what you want for playing real games. Might be good in wac, but not in games. >2. For check, mating, promotion, capture-near-king: set a limit based on the >normal (unextended) depth of the tree under this node. For instance ExtLimit = >Depth/4 allows one extensios every 4 plies. some sort of limit is needed. I allow 1 ply of extension for every one ply of search, until I reach 2*iteration_depth.. then I tone it back 50% to keep it from being unbounded. > >3. Recapture extensions: Don know?! Probably same as in 2? > > >Would this work? And is it optimal? If not, what is? > > >Thanks in advance, >Bas Hamstra. > >PS: Currently I skip nullmove in a node where any extension is added. The idea >is that if there is an extension, there is danger (well, uncertaincy) so it >doesn't seem right to do null then. Reactions? I do nulls _everywhere_ except for where the hash table says "don't" (this from the deep thought extension paper). But I have that 1 ply max extension limit _everywhere_ with no exceptions (note that I extend when I _give_ a check, so that the one-reply extension is done when I get out of check letting both be done.
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.