Author: Robert Hyatt
Date: 12:22:53 09/24/01
Go up one level in this thread
On September 24, 2001 at 14:50:27, Mike S. wrote: >On September 24, 2001 at 11:28:13, Robert Hyatt wrote: > >>(...) >>Again, If you take the same two programs, and vary _only_ the hardware, then >>any difference _must_ be the hardware. SSDF is not passing learning info >>around between the testers. Most programs that I know of (Crafty being the >>one exception) don't have an external portable learning file... > >I think this is one of the issues, where big improvements could be made... I >imagine learning data pools for each program, with users being able to upload >their individual new learning data, and download huge amounts of that. Maybe >those are organized by depth it was calculated at, openings or whatever. Or >users can exchange that data with each other, and add it to their own files, >etc. > >Probably this idea is limited by the quick release of new versions also among >the pros (once a year or so), so that the old lerning data would be outdated >soon and not be useful anymore, when a program has been changed. > >But it's still interesting, also the update cycles seem to take more time now >than in previous years. > >Does the Crafty learning data contain only positions and their evaluations, or >is there a kind of performance stamp added? I assume the search depth is >included, at which the info was produced? I'm sometimes not quite sure with >other programs, if they avoid using data from i.e. blitz/9 plies also when they >play at 40/2h and could probably reach 3 plies more to reconsider. Usually, the >docs don't go into details that deep. But I think that problem should be solved >(?). For book learning, the learning data considers the positions (opening sequence of moves, actually), plus the speed of the search (deeper searches yield more trusted learning values), plus the rating of the engine and the opponent, so that games against stronger opponents influences the learning stuff more than games against weak opponents. (unless the weak opponent does well, then the learned values are very trusted.) > >Regards, >Mike Scheidl
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.