Re: [gnugo-devel] 9x9 go

[Evan Berggren Daniel]

On Sun, 7 Dec 2003, Heikki Levanto wrote:

> On Sat, Dec 06, 2003 at 08:25:02PM -0500, Evan Berggren Daniel wrote:
> > [...] I hope to add some modifications to my metamachine code
> > to auto-generate testcases
>
> Would it make sense to build a testcasefinder directly into GnuGo?
> If the estimated score at move N is considerably less than at move N-1,
> something bad must have happened! If a group changes from
> unconditionally alive to unconditionally dead in one move, something is
> obviously wrong. In such cases, the program could write out what it
> found strange, probably as a commented test case file. Of course human
> evaluation would be needed to decide if those test cases are relevant.

We have something similar in the matcher_check script.  The script could
use improvement, but the basic idea is to replay a game (or play two gtp
engines against each other) and watch the output of the dragon_status
command for illegal transitions (dead -> alive, alive -> dead, uncertain
-> dead).  It has a (very) rudimentary automatic testcase generator.

It is the source of the auto_gen tests, but those are kinda old;
rerunning it with 3.5.2 might be a decent idea.  The script could also use
improvement if anyone wants to work on it.

The other reason to use my metamachine code is that at present it has a
better score estimation than gnugo (usually), and it's relatively easy to
manually add branches for consideration, and it can load and use other
sgfs, and I have a modest sized game tree built for it already.

But in short, yes, your wild idea is a good one :)

Evan Daniel