Re: [gnugo-devel] cosmic patch

[Arend Bayer]

On Sun, 10 Aug 2003 address@hidden wrote:
> Nicolet wrote:
>
> > It seems that the fail/passed status of that test in essentialy
> > random. I think we are in front of a weird cached memory issue
> > there, or maybe hitting a floating point precision limit.
> >
> > With the *same* executable (current cvs, non cosmic, without
> > recompiling), I get different results when I issue the two
> > identical commands in a row in a console window :
>
> The regressions are deterministic. However if there are two

Or they used to be. I can reproduce what Nicolet described by running:
gq -l games/nngs/Lazarus-gnugo-3.1.31-200204140254.sgf -L 90 -r 6 -t -w

If I change the "6" to "5", GNU Go gets a different status for the P10
group.
I have tracked this down to a side effect of different hash collisions
with the killer move heuristic by Evan. (Eventually, find_defense()
returns a different move to owl.)

> It may not be important to figure out the exact mechanism. Perhaps
> we should not worry about this.

We should fix the random seed that computes the Zobrist hash numbers.

But this has all nothing to do with cosmic_5_1.2a.

Arend