Author: Robert Hyatt
Date: 21:11:44 07/27/01
Go up one level in this thread
On July 27, 2001 at 11:18:21, Georg v. Zimmermann wrote: >I came up with a "new" idea. > >One main problem I see with self learning is that you need lots and lots of >tests. Lets say you need 500 games to know if the adjustment really improved >your eval, and you have (only !) 8 changing parameters ... > >So you could try to write endgame functions using self learning, since the >changes can be tested much faster (you can test the change on 500 positions >where the result is known through the Endgame-Tablebase). > >What do you think ? > >Regards, >Georg It isn't so easy. IE in KR vs KQ, it is really easy for the KQ to win without tablebases. But many of the defenses are really odd, and trying to tune an evaluation to mimic the tablebases is very difficult, because even a human has a hard time trying to study a tablebase to learn how to play that particular ending properly.
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.