[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] GNUgo on NNGS
From: |
Paul Pogonyshev |
Subject: |
Re: [gnugo-devel] GNUgo on NNGS |
Date: |
Sun, 5 Oct 2003 02:10:54 +0000 |
User-agent: |
KMail/1.5.9 |
Dan wrote:
> > Hi !
> > What explain that GNUgo lost almost 1 Kyu on NNGS ?
> > md
>
> Suppose we try turning the Cosmic option off and see if it
> goes back up?
>
> When the Cosmic patch was made default in 3.5.1, we had
> 134 unexpected FAILS and 39 unexpected PASS in the regressions,
> a net loss of 95. See regression/BREAKAGE for a list of these
> changes. Also games that I've watched on NNGS have seemed to
> play worse than 3.4. These days most games are depressing to
> watch, and I did not have that feeling during the weeks
> leading up to 3.4.
I agree that the cosmic patch is not very useful (yet). Frankly
speaking, i don't like the basic idea.
As far as i understand, the patch weights quite well tuned territorial
valuations. The effect is probably an emulation of strategical value/
moyo value.
What i don't like in this approach is the distortion in more or less
precise territorial valuation. I'd rather prefer a separate strategy
module that would use some heuristics (maybe neural networks) to
e.g. determine weak dragons which are useful to be pushed and chased
around to gain influence/territory.
However, i'm not absolutely against the patch. If it were in better
shape, i'd have nothing against it as a temporarily solution.
I've seen a game on NNGS recently where GNU Go created a huge
surrounded area which was obviously invadeable. I guess this was an
effect of cosmic patch -- overevaluation of moyo/center territory.
The game ended in 100+ point loss.
Paul