I am not sure what you mean by "sigmoid has changed" but any change to the
way it is evaluated means the net will suffer - it is not recommended
without re-training.

-Joseph

On Tue, Sep 1, 2009 at 7:24 AM, Michael Petch <[email protected]>wrote:

>
>
> On 31/08/09 1:09 PM, "[email protected]"
> <[email protected]> wrote:
>
> > Hi Michael
> > I'm almost certain the problem arises when at the 2-ply stage gnubg
> thinks the
> > move is an error it simply references the 0-ply evals but reports them as
> > 4-ply.
> >
> > This would explain why it takes longer with pruning on than off.
> >
> > The problem doesn't occur when the move played is the best move at 2-ply.
> >
> > Gnubg then reports the correct 4-ply evals.
> >
> > I've checked quite a few positions and they all confirm my findings.
> >
> I am CC'ing this back to the list to provide others that information.
>
> > On a separate note I'm using build Aug 3 2007 because I can't get the
> latest
> > build to stop crashing. I noticed a lot of the times the evals will
> differ but
> > I've never seen more than 0.001. Do you know why that is?
> >
>
> With regards to the crashing. What happens if you remove gnubgautorc (make
> a
> backup somewhere else) from the .gnubg directory (Its a subfolder of your
> home directory), and then reenter all of your settings. Does it crash? If
> it
> continues to crash, what are you doing when it does crash?
>
> As for evals being different, yes there is a reason. Since 2007 we have
> added two things. SSE vector support. This changed the evals slightly. More
> recently we added SSE2 vector support and a new sigmoid function in the
> neural net. The evals will be slightly different yet again. Since 2007
> there
> have been a number of bug fixes, some of which may alter rollout results as
> well.
>
> You can not rely on evaluations to always be the same from release to
> release. When you install Gnugb there is a file installed called Changelog.
> It contains a list of the bigger changes that have been committed (and the
> date).
>
> Michael
>
>
>
>
> _______________________________________________
> Bug-gnubg mailing list
> [email protected]
> http://lists.gnu.org/mailman/listinfo/bug-gnubg
>
_______________________________________________
Bug-gnubg mailing list
[email protected]
http://lists.gnu.org/mailman/listinfo/bug-gnubg

Reply via email to