Re: [gnugo-devel] Owl persistent cache

[Inge Wallin]

Arend wrote:
> On Thu, 28 Feb 2002, Daniel Bump wrote:
> 
> >
> > Inge wrote:
> >
> > > After computing the result, it recalculated it again using the
> > > contents of the hash table.  Thus many blind alleys were never
> > > searched, and the active area turned out much smaller, sometimes only
> > > half as big as before.
> >
> > This seems drastic since then the code is run twice.
> This method would look doable if the second run was some magnitudes faster.
> It would be automatically faster because in WIN situations, we know the 
> correct
> answer at once. Unfortunately, in LOOSE situations, we would have to go
> through the move generation to produce the moves tested, which is the
> most expensive part and which requires all the other expensive work of
> updating the owl data.

You have a good point here, but remember that most sitations, at least
open ones, are WINs.  This is because GNU Go doesn't yet try chancy
invasions, but extends from safe positions.  The losses appear mostly
in closed positions.  Maybe we should only try this if we have a win
to start with.

        -Inge