Re: [gnugo-devel] reinstating value(100) owl defense moves?

[Evan Daniel]

On Thu, 2 Sep 2004 00:00:34 +0200 (CEST), Arend Bayer
<address@hidden> wrote:

> That is already a good example for caution required:
> 
> |......
> |.XX.X.
> |.OX...
> |O.OXX.
> |..O..X
> |......
> +------
> 
> Black at 2-2 kills this.

You're right.  I hadn't looked at things very carefully before trying
the patch as a first try.  Clearly more caution and review would be
needed for a full version.

> I am a bit unsure. My fear is that we would have to add some 100
> patterns to see some 1% decrease in owl nodes overall in real life. Given
> that it also adds to the time spent in matchpat.c(), and that it adds
> complexity in the code.
> 
> Maybe you could try out how much your handful of patterns saves when
> replaying "real life" games. Then we may have a guess how much saving
> one could get with, say, some 50 patterns.

I'll try that in the next day or two.

My thinking is that for potentially deep reading situations with well
understood answers, no good go player actually reads out the answer
every game.  It may well be cheap enough to do so, but it seems
unlikely to me that we can come up with a solution that gets things
like tripod, L group, comb formation, etc correct even at high depth
without a large performance penalty elsewhere.  I also think that this
will become more important later as gnugo gets stronger and does more
generalized and deeper reading, not less.

It is also entirely possible that we need to do this in a completely
different manner, and maybe doing this as a first pass is not worth
it.  Something automated would seem a lot less error-prone.  At the
very least, some automated checking would be in order.

Evan