Re: [Bug-gnubg] Training Bots At Match Scores

[Joern Thyssen]

On Fri, Jan 24, 2003 at 10:40:51AM -0000, Ian Shaw wrote
> > From: Øystein O Johansen
> > Sent: 23 January 2003 14:32
> 
> > 
> > I would love to see how a special trained DMP network 
> > performs against a
> > traditional net. Will it be much hacking into the GNU code to 
> > train a DMP
> > net? NUM_OUTPUTS set to 1. Will it be much more?
[snip]

I'm sure specially trained nets will perform better a given match
scores. For me the problem is that the approach is not general, e.g.,
if we had a neural net for gammon price 0 and one for gammon price 0.5
(which we have now), what do we do for intermediate scores?

What net should be used in the first game of a 17 point matche with the
cube on 16? This can probably be solved by some suitable interpolation
between the DMP net and the MG net. 

However, what do we do at the match score -1,-2 where one player has
gammon price 0 and the other 1? We probably also need a net (or rather:
two nets since the score is not symmetric) for this score too.

An alternative solution is to use the gammon price as an input to the
neural net, but that could require more hidden nodes in order to "Store"
the information. Also, this will make training harder since we now would
need to train on "many positions" times "many gammon prices".

Jørn