[Bug-gnubg] Holding game improvement?

[Øystein Johansen]

Hi all!

At GammOnLine there's several discussions of GNU Backgammon, and in one 
of them Neil suggests that GNU needs some improvement in holding game 
evaluations. I actually think he has a point.

Quote:
"Holding games are very very common in BG and if GNU were a bit better 
with the cube in them I feel it definately would be the best bot out there."

Just to start a discussion I'm mailing some of my opinions of how this 
can be done. As I see it there's two possible techniques to get better 
holding game evaluations.

1. Train the current net further.

2. Add a new neural net for handling holding games only.

The first method is very simple in theory. Can we add some holding game 
positions to a training database and run a couple of epochs? The 
problems will be to verify that this new net wil play stonger than the 
previous net. We may risk that the net forgets some of the skill it has 
learned previous training sessions. Will the database of reference 
positions report this.

The second method is a bit trickier. First we need a definition of 
holding game. Then we must train this net, and then we may have the 
problem of the discontinuities of the nets.

This is just some lines to start a discussion and to hear the different 
opinions here at the list. Joseph?

-Øystein