| |
BGonline.org Forums
AlphaZero pseudocode now available
Posted By: Timothy Chow In Response To: AlphaZero pseudocode now available (Frank Berger)
Date: Monday, 10 December 2018, at 12:03 a.m.
While processing power certainly makes a difference, the crucial difference from Tesauro's work, as I see it, is that temporal difference learning has a tough time learning things such as how to roll a prime all the way around the board, whereas the new techniques for training multi-layer neural nets, combined with MCTS, seem to be so powerful that I expect them to have no trouble with that. Granted, in some sense nobody really understands why deep learning works, so any prediction about its competence in some untested area must be somewhat tentative. But given its amazing success in go, shogi, and chess without any domain-specific expert knowledge, I don't see why we should be skeptical about backgammon.
I'm rather startled to see you say that introducing dice doesn't suit MCTS. My intuition is exactly the opposition. Dice are a very natural fit for MCTS. The MC part of MCTS is throwing dice, after all! Until very recently, the conventional wisdom was that what you call a "straight alpha-beta game" such as chess was not suited to MCTS, because (so the conventional wisdom said) chess is too tactical. And in fact, the most visible weaknesses of AlphaZero Chess are precisely in the tactical arena. If you want to find an obscure mate-in-30 in the endgame, Stockfish is still by far a better bet than AlphaZero Chess. Nevertheless, AlphaZero is able to handle tactics pretty well anyway. To me that suggests that AlphaZero backgammon ought to be able to reproduce not only the positional sense of standard backgammon bots, but also integrate that ability with tactics such as prime-rolling, thereby eliminating the last Achilles heel of standard bots.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.