Re: [gnugo-devel] inconsistnies in owl_attack and owl_does_attack

[Gunnar Farneback]

Evan wrote:
> OK, I've been looking at and thinking about the results from the automated
> regression tests.  Basically, under some cases we get illegal switches in
> the result of the matcher_status (now dragon_status?) function.  This is a
> result, I believe, of moves which gnugo can tell are attacks on a dragon
> but does not recognize as such when listing attacks.  (ie, not included in
> the output of owl_attack, but owl_does_attack returns true.)

If this is the case, we are in a reasonably good position.

> Another option that doesn't fix anything but will probably improve play
> would be to run owl_does_attack on all the top moves proposed, and make
> sure that all attacks are being counted in the final valuation.  This way,
> if, say, the third ranked move has an attack that isn't found, with a high
> enough value to be worthwhile, gnugo will find it and play that move
> instead of the previously ranked top move.

This sounds very unlikely to be effective. If a dragon is thought to
be alive, a missing owl attack move is usually not valued very high,
if evaluated at all.

> Anyone else have any thoughts?  I propose to begin working on this, but
> probably won't get much done for a week or two...  school's starting back
> up.

Improving and tuning the owl code is the way to go. A step on the way
might be to systematically find out where it most often goes wrong.

/Gunnar