[Bug-gnubg] Re: Bug-gnubg Digest, Vol 66, Issue 12

[Frank Berger]

>
Hi Misja,

> I'm purchasing a new smartphone at the moment and definitely want to  
> have
> bgmobile on it too..
> I was wondering how the playing strength of bgmobile compares to the  
> desktop
> version? I mean, does it use the same engine or did you have to make  
> some
> concessions to optimize cpu or memory?

With the limited resource of a mobile there were some tradeoffs  
necessary, so the desktop version plays in another league (2035  
according Gnu and the forthcoming 2.6 will be stronger):

- The principal algorithm is the same, but the net is much smaller (40  
neurons in the hidden layer instead of 200)
but the BGBlitz net (and GnuBG most probably too) is pretty robust  
when you make the net smaller.

- I use only 1-ply (i.e. 0-ply in GnuBG speak). I guess that a lot of  
phones are capable of playing 2-ply (1-ply) and I will try this out,  
probably after BGB 2.6 is out. This will gain 30-40 rating points  
approx I guess.

- older phones don't have floating point and with the newer (CLDC 1.1)  
I don't know how fast they are, as floating point stuff isn't the  
typical domain of a mobile :)) So we used 32-Bit fixed point  
arithemtik. That leads sometimes to numerical quirks. With bgb2go 1.5  
it is now much better, but I'm not sure it won't occure anymore.


I made recently a couple of 3 point matches and put it through GnuBG  
analyser. The resulting Fibs rating was between 1880 and 1890 IIRC. I  
don't remember the error rate anymore (5 or so?) and there was one  
single cube decision that was so weired, it has cost a full point  
error rate and it was probably a bug (or a numerical quirk). I'll  
investigate that later. So the bottom line is: No comparison to a  
desktop, but a pretty good player that is always available.


I hope that answers your question, if not, just drop me a mail

ciao
Frank