Author: Robert Hyatt
Date: 21:34:31 03/01/05
Go up one level in this thread
On March 01, 2005 at 17:23:20, Peter Skinner wrote: >On March 01, 2005 at 12:10:33, Robert Hyatt wrote: > >>No. Position learning is pretty "opponent-specific". And it stored exact >>positional scores. If I make any change in the evaluation, then those scores >>are wrong and can't be used. SO every version change needs to delete >>position.bin at the very least... Book learning is also based on positional >>scores, and are also opponent-specific. If you let it go too long without >>clearing, it will eventually lose a game with most every opening, obviously, and >>flag them all as bad, when for some of the games, it wasn't the opening that was >>inferior, the opponent simply played better. > >Wouldn't it be better to make the learning global instead of opponent specific? > >With nearly everyone using Shredder online, it would be best to learn _those_ >lines instead of just for this person or that person. When I said "opponent specific" I mean that what it does is to prevent one opponent from finding a book exit point where he can beat the program over and over. Two different opponents rarely play the same openings so the position learning is totally wasted there... > >Also most titled players use the latest theory in their openings. Learning that >would as well be beneficial. > >Peter The math doesn't work. A good book could be made from a PGN file with 100,000,000 moves. Position.bin comes from searches done _below_ some of those book lines. The likelihood of seeing a position a second time from a different opponent is nearly zero.
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.