[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] 9 stone games
From: |
bump |
Subject: |
Re: [gnugo-devel] 9 stone games |
Date: |
Fri, 27 Sep 2002 08:40:13 -0700 |
> Dan wrote:
> > Two elements of the patch may need discussion. I raised
> > the value of j patterns to 27. This is a guess and maybe
> > it's too high.
>
> I'm frequently finding that the current 24 is too high. There are lots
> of j patterns which simply aren't worth that much.
Then we should not use that part of the patch.
> Yes, it's very likely harmful. There is fine tuning (the string
> connection bonus) which for late endgame moves is significant down to
> differences of about 0.02.
Right, in the endgame it could be problematical.
> A better solution if the randomness needs to be increased for specific
> moves would be to introduce a mechanism for this purpose. I'm thinking
> along the line of having a randomness_scaling field in the move data
> which initializes to 1 and can be modified (increased) from patterns
> through a helper or action line.
A very simple solution would be to supply increased randomness
just for Fj and Ft patterns. You are right in that the endgame
> Btw, have you run the regressions for this patch?
They're running now. For the first batch, and up to heikki.tst
there were no changes. But I think I should revise the patch
taking your comments into account so I'll stop them at this
point.
Dan