Author: James Swafford
Date: 20:24:24 02/04/05
Go up one level in this thread
On February 04, 2005 at 22:14:44, Dann Corbit wrote: >On February 04, 2005 at 22:08:32, James Swafford wrote: > >>On February 04, 2005 at 20:26:33, Peter Skinner wrote: >> >>>Hello everyone, >>> >>>We have a new entrant: >>> >>>Phophet >>> >> >>Prophet > >I am curious about how much automated learning made it into the current version. >Do you have an information page about it? None. I haven't done any tdleaf 'runs' for a while. It's been over a year; maybe two. The "learned" version was never able to beat the same version of the program with hand tuned weights. It did manage to come up with reasonable piece values, but some of the piece-square tables were pretty funky looking. I think the problem came down to: 1. dramatic changes in score due to tactical defeats (I should probably ignore large temporal differences) 2. too few eval parameters (it was pretty obvious that the pawn piece square tables were compensating for lack of king safety) Or, more simply- a crappy search with a simple eval. (Or was it a simple search with a crappy eval?) In either case, the situation has improved somewhat, but I haven't repeated the experiment. I might as well, as I have lots of CPU time but little programming time. :) Sorry, no web page. I'll try to soon, but my free time is pretty limited these days. -- James
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.