[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] fuseki db again
From: |
Douglas Ridgway |
Subject: |
Re: [gnugo-devel] fuseki db again |
Date: |
Sat, 9 Apr 2005 08:48:26 -0600 (MDT) |
On Fri, 8 Apr 2005, David G Doshay wrote:
> On 8, Apr 2005, at 10:01 AM, Douglas Ridgway wrote:
> > Would it be time to bring up the fuseki db again? ...
>
> I think the right thing to do is to see over a large number of
> games how much better the new larger database does v.s.
> the old database.
The question here is what metric for "better" is. I wouldn't expect to see
much of a difference in results, because my guess is that most games are
decided by mistakes later in the game. The reasons for a bigger db are:
1. Variability. With the existing db, self-play starts repeating lots of
games after only a couple hundred or so.
2. Style. If you play dan-level openings, the new db will respond
likewise for a few moves longer.
3. Minor theoretical improvements due to cleaner game set, postponing
the beginning of mistakes, and the high variability making it difficult to
tune against GNU Go.
4. Basis for tuning. If GNU Go does badly after a particular fuseki,
it's either a bug to be rectified or a reason to avoid that fuseki, or
both. Avoiding that fuseki will reduce variability, so it's important to
start as wide as possible.
I'm not saying I'd say no to a couple hundred games against MF, but I'd
run them through extract_fuseki and look for tuning opportunities. And of
course, if there were a difference in strength between the old and new db,
that would be important to know.
doug.