Re: [gnugo-devel] large_scale_5_1.1b

[bump]

Arend wrote:

> > disabled by default. So there would be no harm in adding
> > this patch to the CVS. We need to decide whether we
> > are still interested in this patch given that the breakage
> > is now about 50/50.
> 
> I still think that this patch can be very useful. But I don't see much
> point in adding it unless someone wants to takesup the work to make it work
> well.
> --experimental-owl is another case of useful code that is waiting for
> someone to put it in shape.

I've looked at the failures enough to be convinced that it is
already a good patch, though whether or not it is worth the
performance price, I'm unsure. In other words, it already
works well enough that it makes the engine stronger but
slower.

Many of the failures are actually playable moves.  One very
bad one is nngs:700 where the patch likes (yuck!) K16.  This
is a case of uncovering an underlying problem with the owl
code, however. And this is exceptional: many of the fails
are not too bad, and many of the unexpected passes are
rather good.

I favor adding it to the CVS. I'll want to look at it a bit
more, though. 

I'm not always able to understand the underlying mechanism of
the failures. For example in arend:9, without the patch, S17
is accepted as an owl attack, with the message

Move at S17 upgraded to owl attack on S16 (WIN)

However after the patch, the move is not upgraded, and I
have yet been unable to figure out the reason for this,
since the large scale code executes but does not add any
owl reasons.

Dan