Author: Stefano Gemma
Date: 06:50:12 04/28/04
Go up one level in this thread
On April 27, 2004 at 16:33:56, Pat King wrote: >On April 27, 2004 at 05:02:11, Stefano Gemma wrote: > >>Yes, i already use move ordering and iterative deepening. In my old programs i >>had used killer moves. In the new one, i'm trying something new (genetical >>algorithm etc). > >I've tried genetic algorithms before. Depending on how you implement them, >convergence will be extremely slow to non-existent! > >The obvious choice of keeping winners and throwing out losers proves absolutely >NOTHING after one game, and there have been many discussions on here that place >the number of games needed to detect a clear difference between programs at >5-30. When dealing with the very small differences between two sets of weights >in the same program, I don't think there's any upper limit to the games you >might need to draw a correct conclusion, and so instead of "evolving", you just >end up with a bunch of more or less random weights. True, if you apply the GA to the wole game. I'm trying to apply GA directly in my version of alfa-beta, at node level, not at game level and at run-time. The idea is to have some "element" that could be detected by the evaluation function (in the future, directly by a neural network) and then to give a bonus to some move (or sequence of move) for any different group of elements. The algorithm themself is not so easy to explain... and still does'nt works ;-). I need to speed up alfa-beta, before to test exaustly the GA. Because of the poor evaluation fuction that i use now, maybe alfa-beta doesn't works so well as it could be. Ciao!!! Stefano Gemma
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.