Author: Simon Waters
Date: 16:32:13 05/25/00
Go up one level in this thread
On May 22, 2000 at 19:55:54, Peter McKenzie wrote: >In any case, the main point of my post was related to how such tuning (whether >it be related to material or positional weightings) can be accurately tested. >So far nobody has commented anything about the validity or not of selfplay >testing! Hmm - well I tried a similar thing with simpler games and genetic type approaches. Here we needed some weighting to optimise play at a game - so I created a scheme to modify the weightings of winning programs to create new programs, and made them self play. The new generations would learn to beat the old ones, their off spring would beat them, and so on - however after a few generations it was clear the strategies were going in circles, like a never ending games of paper, scissors, stone. The lesson here is that most of evolution is not a struggle upwards - but running to stay where we are in the food change - that sounds like life to me. However whether this is applicable to your situation depends on what you are trying to tune, and how that feeds back. If it is just one value - like material value of a pawn - then I would expect it to be reasonably effective. Since many are hand tuned, and systemmatic approach to tuning will probably yield some gains - but don't expect large changes. Remember to pick a constant unit to base tuning around - most people use the material value of the pawn 8-) Gnuchess seems to use a test suite - followed by some bouts on the ICS. There are some classic papers on this kind of tuning problem and lots of research.
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.