Re: [Bug-gnubg] Is it time for Gnubg 0.15? New positions for training database

[Ian Shaw]

Albert Silver wrote:
> I meant the training positions. Would it help to get you (and anyone working
> on the neural nets) more positions? It's not easy to get a big number of
> them, but it is relatively easy to get some positions though. 
>
> An easy way to do so, is to have Gnubg play against itself, and have Snowie
> 4 analyze the matches. You then take every move Snowie says is an error, and
> check that it is indeed an error. These errors would then be the base of
> positions.
>   
This sounds like the start of the next theme I want to explore (so I've 
changed the Subject line).

The simplest way do generate new training positions would be to continue 
with the current system, which is to take positions where gnubg 0-ply 
disagrees with 2-ply. It might also be interesting to throw 1-ply or 
3-ply disagreements into the mix, since they can often be more extreme 
than even-ply comparisons.

Using positions that Snowie disagrees with would certainly be an 
interesting set of data. It would give an independent viewpoint and 
highlight some positions where gnubg is set in its ways. Can we get the 
export format we need. Would it be easier to simply let Snowie play 
itself, export that and analyses with gnubg? If you have Snowie and are 
prepared to do this, then that would be cool. The same would apply for 
BgBlitz. Is using another bot "playing fair"?

Gnubg is known to be weak in its evaluation of holding games, 
consistently overvaluing the holder's equity. Neil Kazaross thinks that 
3-ply produces the most accurate evaluations. Is there a way of 
automatically generating such positions to add to the database? I'm sure 
that you, Joseph, have a tool to automate your 0/2-ply exports. Can it 
be adapted to looks for certain position characteristics?

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.

Again, if I'm not going to roll out the position database, I have some 
PC cycles to put to use. And again, I'll need some help getting the 
tools going.

-- Ian