Re: [gnugo-devel] Owl node limits

[Arend Bayer]

> > Also, there are a number of owl problems that are tough to tune, because
> > it's not obvious how to cut the nodes to under 1000.  Increasing the node
> > limit to 2000 may make tuning of some of these problems easier.
> >
> > That is, there may be obvious tuning that can take place btw the leaves
> > at depths 1000 & 2000.
>
> I guess we need to do some timing tests to see what the impact
> of increasing the owl node limit to 1500 or more would be.

I also did some experiments around versions 3.1.32-33 with a gtp match
between two identical GNU Go versions, one playing with default owl node
limit, the other with 2500. The results were promising but inconclusive,
unfortunately I've lost them. The winning percentage of the 2500 owl
nodes limit version was higher than 60%, but it was playing black and
Dan's gtp matches at that time showed a typical winning percentage of
58% for black with identical engines. Also, it was only around 40 games.

The time penalty varied a lot between games, but on average it was only
10%.

I'd also assume that an increased owl node limit makes sense. 3.2 has
lots more owl patterns than 3.0, so the branching factor probably has
increased, making node limit hits more likely. Hitting owl node limit is
probably the worst case regarding the time cost/gained information
ratio.

Arend


/---------------------------------------------------------------------\
|                             Arend Bayer                             |
|                          Flodelingsweg 27a                          |
|                             53121 Bonn                              |
|                            0228-9813803                             |
|                         address@hidden                          |
\---------------------------------------------------------------------/