> I suspect that for very long time controls we would be better off > turning UCT (with, say 10K playouts) into an evaluation function and > then using alpha-beta on top of it. That is an interesting idea. Usually, when you have to resort to things like this it means that we need a new way of thinking about things. It will turn out at some point that there is some unifying abstraction that we are missing. When it's discovered we will slap our heads and wonder why we didn't think of it sooner.
Perhaps we should be using alpha-beta all along. Alpha-beta with hash tables starts to look more like a best-first search. The hash tables keep the tree in memory. Old fashion alpha-beta (without transposition tables) is rather different. MTF(f) looks more like what we are doing. Instead of iterating in the traditional way you may do a 7 ply search several times - bringing back information each time. So perhaps there is an interesting formulation of tree search that will essentially be doing what you are suggesting. The whole business is rather ugly in my opinon, despite the fact that it works so well. All we are doing is controlling the play-outs. We are starting at the current position and playing to the end of the game a bunch of times. The only reason we do the "play-outs" is due to memory (and perhaps performance) constraints. So we have this artificial wall between play-outs and tree. In fact most of the improvements in MC has been trying to make the play-outs act more like the tree! - Don > > Álvaro. > > ------------------------------------------------------------------------ > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
