Re: [gnugo-devel] Opponent follow-up value and group connection

[Evan Berggren Daniel]

On Mon, 12 May 2003, SP LEE wrote:

> You are right. I was afraid that this solution was too simple as well. Is
> there a good way to tune the valuation system? And is the value equal to
> point?

Unfortunately, there is not a good way to improve it generally that we
know of.  The value of a move is supposed to be equal to the point value
by deiri counting (ie picking up 10 stones is a 20 point move).

If you are looking for broad general approaches, there are some things you
might try.  One thing is that sometimes a simple capture or rescue is
worth more than is immediately obvious because rescuing the group would
leave weaknesses in the opponent's position.  I do not believe we take
this into account in any way; better finding of reverse followup values
might help here.  Another thing isthat frequently large strategic effects
on multiple groups are wrong.  It is unlikely that a single move could
have effects worth 5-10 points each on multiple different groups; some
careful method of limiting this might help.  We currently value joseki
moves fairly highly (35 points).  This is likely too high in many cases,
and you might try simply lowering it by a few points.

Unfortunately, this is not an easy problem to solve.  The best approach
might be to work at expanding the regression test suite to get better
coverage of the problem, and then looking for common themes in the
positions GNU Go has trouble with.

If you make any of these changes, further testing beyond the regression
test suite is likely required.  You could do some such testing by running
long series between patched and unpatched versions, or by using the
matcher_check script with the --stable option to look for differences in
move choice between the two versions.

Good luck

Evan Daniel