BGonline.org Forums

Training bots on alternative starting positions

Posted By: Timothy Chow
Date: Thursday, 26 November 2009, at 4:56 p.m.

Has anyone tried to train a neural network specifically for Nackgammon, or even Snake?

Even for players who are interested only in standard backgammon, I think it might be valuable to see what happens if these nonstandard starting positions are used to train a neural network. One might hope, for example, that a network that has been trained on Snake as well as on the standard starting position would have a much better understanding of deep backgames, which is one of the Achilles' heels of current bots.

Although changing the starting position changes the character of the game, it is still backgammon, and therefore in some sense still the "same game." If we're looking for ways to take the next leap forward in bot technology, then expanding the pool of positions that the neural nets get to train on would seem to be a promising tack.

If Snake seems too radical, one could start with Nackgammon. Train a bot with Nackgammon, and then roll out the standard openings. Does it give the same answers as the current bots do? If so, we've learned something. If not, then try Snake and see if the answer is still the same.

Messages In This Thread

• Training bots on alternative starting positions
  Timothy Chow -- Thursday, 26 November 2009, at 4:56 p.m.
  • Training bots on alternative starting positions
    Frank Berger -- Thursday, 26 November 2009, at 8:23 p.m.
    • Training bots on alternative starting positions
      Timothy Chow -- Friday, 27 November 2009, at 3:07 a.m.
      • Training bots on alternative starting positions
        Frank Berger -- Friday, 27 November 2009, at 7:58 p.m.
      • Training bots on alternative starting positions
        playBunny -- Sunday, 29 November 2009, at 11:24 a.m.