> 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. > > Álvaro. > I did do a study once with pure alpha beta where I used play-outs as my evaluation function. Interestingly, what seems to be the case is that as you search deeper, you need less play-outs. You can measure this by plotting different depths and play-out counts and try to maximize the strength vs time ratio.
It also turns out that it is difficult to get a search to go fast if you do not have enough play-outs. Sometimes it was faster to do twice as many play-outs. This is because of move ordering issues. A small number of play-outs really hurts the number of beta cutoffs you can expect to get because of scoring uncertainties. With heavy play-outs, this would likely be less of a factor. - Don _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
