Author: Marc Bourzutschky
Date: 05:12:01 05/23/04
Go up one level in this thread
On May 23, 2004 at 01:38:38, Ed Trice wrote: >I have toyed with the notion of using R+P vs. R+P to seed a game-theoretical >value database of R+P+P vs. R+P, then R+P+P vs. R+P+P. > >While the distance-to-win for conversion would not be know, this would not be >sought. Just store W-L-D values in a highly-compressed RAM buffer. This way, >from a great distance, a program can head for theoretically won positions that >otherwise would be misdiagnosed as leaf nodes. > >There are a few details to work out, but something tells me you can jump right >to the pawn slices in exchange for some nomail depth searching from the first >pass, then the typical 1-ply passes subsequently. The devil is always in the details with these things. The biggest problem is to work out the value of endgames resulting from promotion of one of the pawns, very messy without generating the full subgame databases. There is probably not a huge gain in creating WLD data instead of distance to win data. In distance to win generation, you typically break the problem into successive (N+1) to N ply computations, which takes the same amount of memory as a WLD computation. The only difference is that you need to separately store configurations for each N, rather than combining all N. You have more disk IO, but by compressing each N ply configuration that should be managable. -Marc
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.