Author: Wylie Garvin
Date: 13:05:28 12/12/01
Go up one level in this thread
On December 12, 2001 at 15:41:02, john rice wrote: >Hello, >Since larger hashtables increase playing strength, isn't it safe to assume >increasing the tablebase cache would do the same? With 512m of ram available, I >usually set the hashtable size of Fritz 7 to 128 or 256m, yet the default >tablebase cache is only 2m. With all 3,4,5 and some 6 piece endgame tablebases >it would seem 2m of cache would be a huge slowdown/bottleneck during the >endgame. Any thoughts? >Thanks, >JR Hi John, I am under the (perhaps mistaken?) impression that most programs limit their accesses to the tablebases to some shallow depths. Otherwise they have a tendency to quickly become I/O-bound with a cache as small as 2mb, as you have indicated. Part of the problem is that most programs are not equipped to efficiently recognize database positions at every node in the tree. Solution? They only try to access that cache for a fraction of the nodes in a large tree. I don't think I'm the only one, who thinks that significantly improved play across the middlegame-to-endgame transition would result if these engines could access the databases at every node in the tree. Someday soon (when my exams are over?) I'm hoping to experiment along these lines by generating compressed bitbases for all the 5-piece endgames. These will be substantially smaller (i.e. 4-8 times smaller uncompressed) than the distance-to-mate type tablebases. They will only yield game-theoretic values, but I will use a large cache and interior node recognition techniques developed by Heinz et. al. to get these values everywhere I can. e.g. 192MB hash table and 40MB database cache. I'm fairly sure this has already been done, I've seen bitbases mentioned in connection with some commercial chess program, but I want to do it myself and see what I can make of it. regards, wylie
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.