Re: AW: [gnugo-devel] Symmetry corrections

[Gunnar Farneback]

Arend wrote:
> Sorry, I think I missed what exactly you did. I think it is fine to try
> to tune pattern <> top_pattern comparison to get as many PASSes as
> possible, whereas this does not make sense to me for the move <>
> top_move. It's definitely good to have both comparisons though, as it
> reduces tuning noise.

I agree that ordering with respect to the move coordinate doesn't
really feel good but I think it, at least for the time being, is more
valuable to get rid of owl tuning noise.

Teun wrote:
> 3.3.9 has 23 reorientation inconsistencies.
> Gunnar's patch brings that down to 17.
> Nando's revision brings it back to 23, though
> the inconsistencies are somewhat different from unpatched 3.3.9.

That's more or less what I would've guessed.

> The problem in Nando's patch with regard to reorientation invariance
> is this line I think:
> 
> +               && list->pattern_list[bottom].move < top_move

It absolutely is. The orientation effects is the main reason why
ordering on the move coordinate is unnatural. On the other hand, it
may be just as well that these inconsistencies are made visible
instead of having them pass by chance.

/Gunnar