Re: [gnugo-devel] fast_defense() improvement

[Arend Bayer]

> nando wrote:
> > For Paul: "proving" the safety looks difficult. It would involve reading,
> > but that's precisely what we would like to avoid.
> 
> You misunderstood me.  I meant a formal proof of safety of the proposed
> code.  Or at least an informal one, with diagrams and all.  I'd like to
> see some diagrams explaining how it works and why and in which circumstances
> we need to subtract liberties.

I think this is besides the point. I "proved" it under the simplifying
assumption that there is one move to connection all strings in question
at once. Of course, this is not always the case, but other cases should
be rare, and even then I would make a high bet that GNU Go as the
attacker would give up (due to liberty cutoffs) a couple of moves later.

So it becomes a heuristic: "Assume all attacking moves are ineffective in
the situation described as above with the appropriate total liberty
count." That doesn't sound so different to me to the heuristic used
in attack3 "Assume all self-atari attacking moves are ineffective once we
are past stackp > depth". And many other heuristics in reading.c. Most of
which are worth less than a 5% speed gain.

Arend