Re: [gnugo-devel] reading connections

[cazenave tristan]

new files at:

http://www.ai.univ-paris8.fr/~cazenave/test_golois.tar.gz
http://www.ai.univ-paris8.fr/~cazenave/readconnect.c

> test cases and the ones in regression/reading.tst. In e.g. test case
> 6, N2 also does capture the O2 string, although O3 is a better way of
> doing this because it leaves fewer weaknesses on the outside. The
> tactical reading of GNU Go is not expected to necessarily generate the
> best attack or defense move, just to come up with one that works. Thus
> GNU Go needs additional correct answer alternatives in some of the
> test cases.

In golois, I make golois compute many answers and verify that one of
them
is the one in the test.
I modified test 6 for gnugo.

> 
> How well does Golois do on this test suite? Have you tried Golois on
> GNU Go's reading.tst?
> 

not yet, results to come...

> files LifeDeath0001.sgf to LifeDeath0004.sgf are missing. The problem
> with test case 44 is that C9 is an empty point. This is a (false) eye

corrected.

> The first step might be to revise the code to keep track of the
> effective move so it doesn't just return PASS. Tristan, do you want
> help with developing the code, or is it better if we don't touch it
> more than necessary for the time being?
> 

I have added this functionnality and corrected a bug due to
inappropriate
call to naive ladder in the new readconnect.c
I still have things which are missing but that are very clear in my
mind,
however, any help is appreciated.
My current main problem is debugging readconnect for gnugo.

> The failure in test case 50 seems to be an erroneous test case:
> 
> loadsgf Handtalk980819-1.sgf
> 50 disconnect a2 a2
> #? [1 a2]
> 
> There is no stone at a2 and disconnecting a2 from a2 doesn't make much
> sense. Test cases 57 and 58 seem to have identical problems.
> 

corrected.