RE: [gnugo-devel] More on minimum values

[Portela Fernand]

Arend wrote:

> > Since the worst case is still negative, I'm rejecting my patch in
> > its current form and trying something else.
>
> I think you are setting the bar too high here. (...)

Unfortunately, I think I'm not (in my specific case). I have good
reasons to suspect that the first serie of my twogtp matches was
actually extremely out of bounds. As said, sd for the series where the
patch played white is quite high. Also the patch I tested was a
'light' change. I started testing it with a clearly more aggressive
setting and results were definitely worse. So I'm pretty convinced now
that the correct hypothesis is actually the one I described as the
worst case and I'll have to find something else to solve my problem.

Coming back to Evan's patch, I insist that I'm not saying that the
conclusions are incorrect. As you indicated,

> (...) if you have good reasons beforehand to believe your patch does
> s.th. good (...)

a 100 games serie might be enough if you have good reasons to believe
so. By replying in this thread, I wanted to share some results I got
from a recent and similar experiment, and insist that we should be
careful with twogtp matches results.

So my Evan's patch questioning could be boiled down to following: are
we confident it does something good ?

Actually, I have a couple other questions. Nobody replied to my post in
Evan's previous thread, so I'll repeat the questions here : there's
already a mechanism to differentiate minimum values. Why are we
insisting on adding another one ? If the proposed one is better, why
keep the old one then ? What about simply changing the current one so
that it includes J patterns too ?

/nando