Author: Steven J. Edwards
Date: 06:04:24 06/14/98
Go up one level in this thread
On June 14, 1998 at 03:31:12, jonathan Baxter wrote: >>Please post your results. Others have tried this before, but all >>attempts to come up with a completely accurate (if imprecise) function >>have generated functions more complicated (time wise) than a simple file >>probe. >I find the file probes to be pretty expensive---in KnightCap the disk >seeks >can cost up to 2/3 of the processing time. But I have two reasons for >doing this: speed is one and the other is if you can do it accurately >then you may be able to generalize to more pieces without having to >compute all positions. >I'll certainly post the results. A decent disk cache helps with speed issues. Also, you can store the TB lookups in a transposition table, either a main table or a special table reserved for TB values. And an occasional disk defragmentation can't hurt. Finally, the three man classes are small enough to keep in memory nowadays. One of the first to build an expert system from an endgame database is Ivan Bratko (of Bratko-Kopec test suite fame) who did one for KRK. Jaap van den Herik has also worked quite a bit in this topic. The _ICCA Journal_ indexes will be helpful here. It's a hard problem for the four and five man classes. Humans are still the best pattern recognizers and abstractors for most things, so it was not encouraging to see a recent Kasparov-J. Polgar KRBKR endgame where both players were fumbling move after move (compared to the KRBKR tablebase). Kasparov eventually won, but it was clearly a case of of the old saw "the victor in chess is the player who makes the next to last mistake". -- Steven (sje@mv.mv.com)
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.