[Bug-gnubg] New positions for training database

[Frank Berger]

> > Another source of interesting positions would be to generate games  
> > from
> > the Nackgammon start position. With more chequers back, there will be
> > lots of interesting holding games, as well as a decent share of
> > backgames, which are another challenging area.
>
> Frank Berger once told me that he tried to train a net from
> nackgammon starting position. IIRC the new net turned out to play
> slightly weaker in backgames than the original one, but I don't have
> details about his experiment.
I benched the new NG net against the fully trained BG net.
The result were that the NG-net were wearker, but in NG it did a  
little better. When the new NG matured, it still did better in NG  
than in BG. The funny thing were that the difference in playing  
strength between playing NG and BG for the old net and the new NG-net  
get smaller and smaller. When the new NG net was fully trained the  
difference between the nets were approx. 1point/100 cubeless games.
When I did that in 2002 or so, the HW was much slower, so maybe I try  
it again later.


> Interesting. That's a surprise to me. I would have thought there would
> be benefits. I haven't played much nackgammon myself, so I don't know
> how games evolve. Maybe with both sides having four men back,  
> reasonably
> timed one-sided backgames rarely evolve.
For GNU it may work better, because of the different training methods  
(BGBlitz uses straight TD(lambda)-Training). I would guess it's worth  
a try.


> Another approach might start from prototype holding games and  
> backgames
> such as found in Kit Woolsey's Encyclopaedia. The idea being that  
> these
> are common positions, so it is useful to have plenty of data for
> accurate training.
If you enter the positions, could you please send me a copy?

ciao
Frank