RE: [Bug-gnubg] Using weights files of different sizes

[Ian Shaw]

> -----Original Message-----
> From: Philippe Michel [mailto:address@hidden 
> Sent: 14 December 2010 20:36
> To: Øystein Johansen
> Cc: Ian Shaw; address@hidden
> Subject: Re: [Bug-gnubg] Using weights files of different sizes
> 
> On Tue, 27 Apr 2010, Øystein Johansen wrote:
> 
> > I've uploaded a quick and dirty patch to neuralnetsse.c 
> which should 
> > do the trick.
> > http://www.gnubg.org/media/nnsse.patch
> >
> > A binary weights file with the better net is found here:
> > http://www.gnubg.org/media/alternative-gnubg.wd
> >
> > It should be quite simple. Just patch the code and recompile, and 
> > replace gnubg.wd with the alternative-gnubg.wd Please test 
> it out. If 
> > it's good, we make it "official".
> 
> I have been using code with this patch applied for some time 
> without problems. Mostly with the original weights and 
> occasionally with the alternative above.
> 
> I tried the new net only superficially but didn't notice 
> obvious changes while playing it. Analyzing old matches with 
> both old and new nets found a few differences (and rollouts 
> showed the new net right in the majority of cases), but these 
> were mostly close technical plays like while bearing in 
> against an anchor. Did I miss something ?
> 
> How did you train this new net ? From the same data than the 
> current one, just with a larger hidden layer ?
> 

You didn't miss anything. The new network is indeed trained on the same source 
data but just as a larger hidden layer, as you suggest. The improvements in the 
benchmark were small, so we would not expect to see more than a tiny 
improvement. 

It's good to know that the real-world performance seems to match what we 
expected.

Cheers,
Ian Shaw