[gnugo-devel] big dragons never die

[Arend Bayer]

Well, I guess we have all learned to distrust this proverb (at least as
far as our own dragons are concerned...)

However, GNU Go seems to be a firm believer in this proverb:
A big dragon will pretty much always have a couple of spots where some
shape patterns match. Usually enough to generate 3 shape moves up to
OWL_BRANCH_DEPTH. Add to this 1-2 vital moves at each stage. Now if we
assume that the dragon is really weak, so that it will not get declared
alive in any of the variations, we nowhere get a cutoff in the tree.

This means the tree has a branching factor of 4-5. Even assuming perfect
move ordering, this means that we have around 4^(OWL_BRANCH_DEPTH/2) = 250
nodes at stackp == OWL_BRANCH_DEPTH in the tree. As most of them will
have to be read even deeper, it is clear that 1000 owl nodes isn't
enough to fit this tree in (and raising this number to 2000 won't change
this much).

And if we run out of owl nodes, we use the policy "if in doubt, declare
alive". Hence "Big dragons are never dead with certainty".

I wonder whether there is an easy way to at least reduce this problem.
(Of course we have a couple of policies with respect to incertain owl
results, but they all have their problems.)
S.th. like as "if in doubt, declare critical". Has this been tried
before?

Arend