MCTS seems to be not very useful for some games, but I think it needs a lot more research.
I tried it will chess with really dismal results. Even the mighty Rybka team has a MC mode which is a curiosity, it plays much weaker than the standard program. I have this feeling that you ought to be able to at least approach the same level of strength but it would require a lot of domain specific enhancements appropriate to the particular game. In chess for instance, a 1 ply search with a weak evaluation function and very little effort gives you a better move that a million random simulations, but that is probably not true in GO. So one would have to determine what you can and cannot do well with the different approaches and try to meld them in some way that makes sense. - Don On Tue, 2008-10-28 at 11:13 -0400, Janzert wrote: > David Fotland wrote: > > I've thought about using mcts for Arimaa, and it might be better than > > alpha-beta. Arimaa has a huge branching factor, and since the pieces move > > slowly, human long-term planning works well against computers. Mcts prunes > > better, so it should get deeper PVs. AMAF might help too, since in a long > > plan it is often not critical what order the moves are in. > > > > Someone should give it a try. > > > >> David, is MCTS likely to be useful for Arimaa? > >> > >> Darren > >> > > Jeff Bacher author of Clueless gave UCT w/light playouts a try a few > years ago but without much success. If I'm remembering correctly it > played fairly decent in the end game, but for most of the game was much > to prone to advancing rabbits on the off chance that they could sneak > through. > > I also gave UCT a bit of a try but never got it playing much better than > a 1 ply static evaluation. If I can get my alpha-beta program up to the > same level as yours I would like to go explore MCTS and other search > methods some more. > > There have been a number of new programs in development this year but > I'm not aware of any of the ones that have made it to the stage of > actually playing games being anything other than alpha-beta. > > Janzert > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/
signature.asc
Description: This is a digitally signed message part
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
