Re: [gnugo-devel] pattern-based reading spinoffs (27.7)

[Arend Bayer]

> resolved for quite some time now. There is also the complication that
> this scheme would fail if people disable the caching completely, which
> sometimes is needed for debugging purposes.
Yes, with caching disabled we would still have to store the open results.
Then they would get removed instead of setting the status closed once
the result is known.

> > and add a "dummy" open read results for all caching function at all
> > previous game position. This would take away only a small part of
> > the memory usage of the hash table.
>
> Are those dummies really necessary? To enforce a strict superko test
> (with respect to the entire move history) during reading is much less
> interesting than breaking local cycles.
Well, I guess it would be safe to restrict this to the say 15 latest moves.
15 Hash nodes with ~150 read results take away some 0.0x% of the cache
memory with current implementations. I would like that solution because it
should almost automatically make sure that GNU Go does not violate the
superko rule in actual play. Also, of course it would help saving reading
nodes by possiblye breaking such local cycles earlier if we are already in
a ko fight. But it's not necessary of course.
(I believe that Chess programs, where "superko" situations are far more
relevant, actually use such a scheme.)

> However, this is complementary to an extended komaster scheme, which
> could also prevent some other kinds of ko nonsense.
I agree that it makes sense to go in both directions.

Arend