[Bug-gnubg] Re: This is a classic position to start training GNU

[Øystein Johansen]

Hi,

I see you posting at GammOnLine, and totally agree. We need some more training
on holding game position.

I guess the way to do this is to split the contact net into a holding game,
and train this network to evaluate this position better. (Maybe someone has a
better method?)

OK... if we splitt the contact net into "contact" and "holding", we must have
a clear definition of the holding game class. This may sound simple, but I
don't think it's simple at all. I think the definition should not handle too
many position and not too few. If it handle too many positions it will be slow
to train, and we may have many cases where we get the net discontinuity
problem. A complex code in ClassifyPosition will also slow down the
evaluation. If there is too few position handled by the "holding" net, the
improvment won't be much.

I guess a net for holding game positions should handle about 5-8 % of all
position. I think that's the size range we're talking about. (Just guessing
;-)

I believe the holding game definition should be limited to one side holding
games. (It plays mutal holding games fine as far as I can see.) I also think
we should limit the definition to a minimum of anchor points.

Can you, Neil, or anyone else, take a look at positions where you know GNU
Backgammon evaluates the position poorly, and see if you see anything common
in these positions? Where is the anchor? Where is the furthermost checker? How
many checkers are left on the midpoint? What's the pip lead? Etc. Is there
some kind of system, in the positions with poor evaluation?

-Øystein