Author: Keith Ian Price
Date: 09:58:05 11/27/97
Go up one level in this thread
On November 27, 1997 at 04:38:48, Chris Whittington wrote: > >On November 27, 1997 at 03:38:28, Howard Exner wrote: > >>Does anyone have an example of a series of games where a book >>learning program (MCPro, Crafty, Rebel 9) re-writes a book line >>that initially loses then gradually finds an improved line? > >Yes, but I haven't got the games anymore. Mchess is particularly 'good' >at this, since it plays a limited, but well booked-up, opening set; and >it will tend to repeat specific opening lines again and again, even if >it loses them, if it thinks the score just out of book is good. Examples >are french, urusov gambit typically. If you fix your programs learning >(you being the programmer), to replay won games, then the result can >often be long autoplayer sequences on a specific variation. It is >possible to find 'improved' lines this way. With 'cooperation' between >programmers to lock their programs into specific lines, I guess it would >be possible, with strong programs playing on fast machines, to have >quite an impact on opening theory and to find useful novelties. Maybe it >would be possible to do it via a self-play process as well - new feature >? > >Chris Whittington I noticed this strange behavior of MChess while autoplaying it against CSTal. It would play the same two openings over and over, both of which it would lose. One opening it thought it was half a pawn up when it came out of book, but the other it was -2.37 every time. I finally turned off book learning, and switched to the variety book, and it began to switch openings including some b3 and f4, and began winning most of the games. I can't figure out why it will just not eliminate this opening (-2.37) from its book, or at least switch to another. kp
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.