Re: [gnugo-devel] Joseki implementation question

[Evan Berggren Daniel]

On Thu, 24 Jul 2003, SP LEE wrote:

> Actually I got this question from the regression test nando:212. I didn't
> understand why gnugo doesn't know such simple joseki. Then I see in
> hoshi_keima.db, the sequence for invading the corner is only included for
> ikken basame (not sure if this is the right word). However, this sequence is
> valid for many positions of basame. It seems we need more implementation of
> these.

The basic reason is that not a lot of effort has gone into tuning the
joseki databases.  I made some significant extensions a while ago, but
didn't really tune things much.  I'll look into this one.

Long term, in order to have effective joseki play, we need to do some real
reading to choose the joseki, and also have a way of understanding how
acceptable a tenuki play is at any given point, in a dynamic way.  There
has been talk about doing a localized alpha-beta reader for full board
evalution and having it choose josekis, and I think that will help a lot;
the current method (randomly pick a joseki move) is basically
unacceptable, but properly fixing it will be a lot of work.

Evan Daniel