[Bug-gnubg] Neural Net to play Trictrac

[David Levy]

Hello,

I am a big fan of gnubg -- I use it all the time to analyze my backgammon
matches and it has improved my game tremendously.

I am also very interested in the French game of Trictrac -- a non-racing
game played on a backgammon board. If interested, you can see my Trictrac
Home Page at http://pages.sbcglobal.net/david.levy/trictrac/. I have also
started a Yahoo discussion group at http://groups.yahoo.com/group/trictrac/.

I am working on a computer program to play Trictrac. The program currently
uses a backprop neural net that I can train with supervision. This is
difficult since no one has studied Trictrac for 150 years and the advice in
the literature is quite suspect! I am struggling a great deal in my attempt
to translate the TD(lambda) literature (Sutton et al.) into working computer
code. I was hoping to look through the gnubg source code on the web portal
and see if I could get some pointers.

As I looked through the code, I found a function for training under
supervision in lib/neuralnet.c called NeuralNetTrain(), but didn't see
anything implementing temporal differences.

Am I missing something elswhere in the code?

Much thanks for any direction you can provide.

Best regards,

David Levy