Re: [Bug-gnubg] (no subject)

[Joseph Heled]

Ian Shaw wrote:
> 
> > -----Original Message-----
> > From: Øystein O Johansen [mailto:address@hidden
> >
> > What about the 'database generate' command? Can't that
> > feature be used to
> > generate lots of positions where the 0-ply and 1-ply search
> > differs by x
> > points? It should be quite simple to write a small program
> > which sieves out
> > the positions from the crashed net. Will this create an unrealistic
> > benchmark?
> >
> If you train exclusively on these there is the danger that it might not be
> good at playing the "ordinary" crashed position well. Or will the weights it
> learns from these positions be suitable for all cases?
> 
> > BTW: diff and grep is ported to Windows, even sed and gawk as
> > well. (How do
> > you think I survive this world...)
> >
> I've just downloaded some. I think my brain has atrophied on an exclusive
> Windows diet for the last 4 years.
> 
> Ian
 
I have developed a different set of tools. I hope to fold them back into gnubg
some day, but that is not an easy task. The main point is that I have a TCL
interface for scripting. This has proved a good choice (but if I started today I
would probably choose Python). 


And developing better nets is much harder than I imagined. While the GNUbg
rating has advanced from 1650 to around 2000, each new net was a "leap of
faith". Some of those leaps ended up in a crash, BTW.

-Joseph