|
From: | Nis |
Subject: | Re: [Bug-gnubg] Even, odd and half plies (WAS: New Contact Net Error Rates) |
Date: | Thu, 27 Feb 2003 18:51:14 +0100 |
--On Thursday, February 27, 2003 08:08 +1300 Joseph Heled <address@hidden> wrote:
It seems your idea has lot's of merit. Sorry for being skeptical at first.
Are you saying that I have thought up something which might improve the actual playing strength of gnubg? Woohoo!
107485 contact benchmark positions 0Ply error - 0.01079 (w = 0) 1Ply error - 0.00956 (w = 1) w * ply1 + (1-w) * ply0 w = 0.5 error - 0.009176 w = 0.75 error - 0.009155 w = 0.625 error - 0.009066 w = 0.57 error - 0.009070 w = 0.67 error - 0.009071 Conclusions: - 1ply does play contact positions better after all. - 1.5 ply is better yet. The best w is around 0.6
Just to make sure: The above is what I called 0.5-ply (since it is the average of 0- and 1-ply).
Works for crashed as well, with w somewhere in the 0.4 - 0.5 range.
Could you post the numbers? I am especially interested in the crashed positions - since this is where the largest problems have been reported
...I would also be VERY interested in seeing the results from including 2-ply evaluations, using these two formulas:
w * ply2 + (1-w) * ply1 and w1 * (1 + w2) * ply2 + (1-w1) * ply1 - w2 * w1 * ply0 (w1 and w2 both around 0.5)
So, if we implement such a thing, we will need it to be per class. w may be per class, but that is not very important, as w=0.5 will do quite well.
Please note: The numbers above shows that the _idea_ actually works. I think the best implementation will be using reduced evaluation - probably 50% reduction.
-- Nis Jorgensen Greenpeace Amsterdam
[Prev in Thread] | Current Thread | [Next in Thread] |