| |
BGonline.org Forums
Computer go vs. computer backgammon
Posted By: Maik Stiebler In Response To: Computer go vs. computer backgammon (Timothy Chow)
Date: Saturday, 1 August 2009, at 8:43 a.m.
There have been attempts to create rollout-based players. In fact, ExtremeGammon offers a play setting such that all its moves on based on rollouts, but they are short truncated ones, and move selection is done by the very knowledgable neural nets. More in the spirit of the Monte Carlo methods that are used in go are rollout players that are based on very fast, but rather weak evaluation functions; I was told that pubeval has been used as evaluation function in this context, but I could not retrieve much information about that project.
As I understand, Monte Carlo tree search as it is done by MoGo is effective, because it solves the problem that there are no cheap and good static evaluation functions for Go positions. In contrast, neural nets are exactly that for backgammon positions, and it would seem unnatural to me to ignore their wisdom while deciding on a move. On the other hand, if you have lots of computational power and time at your disposal and want to have a player that is even stronger than a player based on full rollouts at high settings, I can imagine that some combination of knowledge based choice of moves in the rollouts and UTC might be a good idea.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.