Re: [gnugo-devel] Carpenter square test suite

[Gunnar Farneback]

Allan wrote:
> I've made a new test with owl_defend and owl_attack:
> 
> http://www.faldara.co.uk/carpenter2.tst
> 
> However, I get somewhat worse results with this (2 new passes, 6 new
> fails).

In general that can only be determined by checking what the engine
actually does. Try running

gnugo --quiet -l <filename> -L <movenum> -w -t -d0x1000

in a position where you have deviating results and see how the owl
traces relate to the generated move.

> Have I done it wrong?

A quick look at the test file shows that no test has a ko result,
which is somewhat surprising for the carpenter's square. For example
in test 10, black doesn't have enough liberties to live better than
with ko. Furthermore black must also make the first ko threat, which
makes the correct result "3 A8" instead of "1 A8". When the player to
move does not have to to make the first ko threat, the result should
instead be 2. Since the end result of this (optimal) sequence is ko,
presumably all the previous tests should also have ko results.

Another thing that can happen is that the owl code comes up with
another move than the usual one which with respect to life and death
works equally well, but which is somewhat worse with respect to yose
considerations or ko threats. In such cases the alternative move
should also be included in the list of acceptable answers.

/Gunnar