| |
BGonline.org Forums
Very impressive result from DeepMind team - Xavier's take?
Posted By: Maik Stiebler In Response To: Very impressive result from DeepMind team - Xavier's take? (Robert Wachtel)
Date: Tuesday, 19 December 2017, at 8:25 a.m.
The trade-off may not only be the longer training time, but also that, given the limited capacity of existing artificial backgammon brains, the more they know about deep backgames, the more they have to forget about how to play normal backgammon. As David Montgomery observed in 1998:
"Yes, if you are using the same evaluation function everywhere, this will generally be true. For example, in some experiments people tried training not just from the opening position, but also from positions with more checkers back, more potential backgames. In one case they trained from the Nackgammon position. The programs then played better in these sorts of positions, but the overall play decreased."(link)
One remedy is to have separate nets for separate position types. As Xavier informs us here, XG has "many" of those. This I am sure also has its problems, because position types are never entirely separate and there may be ugly discontinuities when going from one net to the other.
Alphazero only has one evaluation net though, and chess according to my understanding is a wee bit more complex than backgammon, so there is hope that a state-of-the-art NN could learn to handle everything well when learning from a balanced diet of self-played backgammon, nackgammon and acey-deucey games.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.