|
From: | Nis |
Subject: | Re: [Bug-gnubg] Even, odd and half plies (WAS: New Contact Net Error Rates) |
Date: | Fri, 28 Feb 2003 10:53:12 +0100 |
--On Thursday, February 27, 2003 10:51 -0800 David Montgomery <address@hidden> wrote:
My understanding of the current implementation is that we sample (striated, as much as possible) from the leaf nodes. So a 100% 1-ply evaluation averages over 36 (computationally optimized to 21, but let's use the 36 basis) best next plays; a 33% 1-ply evaluation averages over 12 best next plays.
Correct.
My understanding of what you are suggesting is that we do this, and combine it in a weighted average with the 0-ply evaluation. Is this right?
It doesn't catch all my ideas, at least. My first thought was to do something like what you suggest - evaluating some of the rolls at 1-ply, some at 0-ply. Then I realized that "evaluating a roll on 0-ply" does not make sense. I found two ways to repair this:
1) Average between 12 rolls at 1-ply and 24 times the static 0-ply eval. 2) Average between 12 rolls at 2-ply and 24 at 1-ply. The second one can only be used for reduced evaluation of2-ply and higher - but it should give superior results compared to any of the other "reduced evaluation" techniques, at no extra cost.
Perhaps we can implement both - using 1) for reduced 1-ply, and 2) for higher plies.
But to know whether this makes sense, we need benchmark results including the 2-ply evaluations ...
-- Nis Jorgensen Amsterdam
[Prev in Thread] | Current Thread | [Next in Thread] |