Re: [gnugo-devel] attack_either patches analysis

[Evan Berggren Daniel]

On Fri, 8 Nov 2002, Arend Bayer wrote:

>
> Dan wrote:
> > About Evan's defend_both patch, and Arend's updating of Arend wrote:
> >
> > > What remains to do is a measurement whether your patches have a negative
> > > performance impact.
>
> > A net increase of about 54 seconds or 4.5%.
>
> This probably means the patch in its current form is not worth its costs.
> However, looking at the code, improvements can be made.
>
> Just looking at attack_either(): We should
>
> 1. just return asuccess when the other string is safe, i.e.
>       if countlib(bstr) >= 5 or maybe 4
> 2. similar if stackp > branch_depth
> 3. really not worry about ko results when you've already called
>       defend_both() (you simply cannot do that when you don't use a
>       komaster scheme).
> 4. try to return as soon as possible (i.e. insert
>       if (defend0 != 0)
>         return defended0;
>       between the two alibs trymove's
> 5. maybe worry about not playing the same move twice (in case they
>       have common liberties).
>
> Evan, do you want to give (some of) it a try?

Yes, I plan to do that.

I'm going to move both functions to the candidate move macros, and maybe
add some depth checks.

Also, I believe the performance hit is closer to 1% overall, if we use a
game replay as the benchmark.  I think the tests are biased against this
sort of patch.  I also think we should have a standard benchmark, either a
game replay or a carefully chosen set of tests.

Thanks for your help on this

Evan Daniel