Re: [gnugo-devel] benchmarks

[Inge Wallin]

Arend wrote
> Here is finally some benchmarking for the break-in code. Overall, it
> costs 20%, which is too much. Especially +43% in the Go4++ game is
> unacceptable, of course.

It is not too much in all situations.  Last year, when I played with
GNU go in the computer olympiad, the thinking time was never an issue.
We played all opponents except the two worst programs on 19x19 using
GMP over a serial cable, and GNU go never used even half of its
alloted time.  On 9x9, the games were over very quickly indeed.

This was not true of the opponents.  Both Go4++ and Go Intellect,
which placed above GNU go in the end, used their time much better.  If
we could come up with ways to use the time to our advantage, I think
that GNU go would benefit much from this.

The breakin code seems to be such an example.  Right now, we can't
control how much time is used dynamically, so we have to use a
specific setting that is set at the start of the game.  Still, the
increase in go knowledge from such a module far outvweighs the time
penalties that it imposes.

> For comparison, I also benchmarked 3.2. Overall, we are a lot faster
> than 3.2, but the variance is extremely high.
> So I think we can spend a little performance on features. This could be
> increasing some depth levels, if we knew which ones are most useful.

Performace optimizations is of course good.  But if I have to chose, I
would not want to exclude such nice features just because it uses some
time.   Besides, when analyzing a game, is persistent caching used to
its full potential?

        -Inge