Author: jonathan Baxter
Date: 17:26:19 10/08/98
Go up one level in this thread
On October 08, 1998 at 08:21:56, Robert Hyatt wrote: >On October 08, 1998 at 04:45:16, blass uri wrote: > >>I think that it is a good idea to have a not symmetric evaluation function. >> >>for example if a program is bad at rook endings it should be against going to it >>with both sides, but the evaluation should be more symmetric when the time >>control is longer because in this case the program has better chances to find by >>search the right moves in positions that it does not understand. >> >>Is there a chess program that uses an evaluation function that is a function of >>the time control? >> >>Uri > >many programs start out like this, but as the evaluation is improved, it is >"turned off". I can't think of anything asymmetric in my evaluation at >present, other than one penalty for completely blocked pawns... One big advantage with a symmetric evaluation function is that you can do parameter learning even if you lose every game. I recently modified my program's eval to be completely symmetric and got very useful learning playing against crafty even though it lost a large portion of the early games. If the eval is asymmetric and the program loses most of its games it will tend to use the asymmetric features to learn that every position is a loss, which isn't very helpful. Jonathan Baxter that "every position is a loss"
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.