On Tue, 2006-10-10 at 10:17 +0200, [EMAIL PROTECTED] wrote: > So indeed choosing the move with highest number of simulations seem a > little > better, whereas it is not statistically very significant (I could try > with > more games, but 800 is already quite a lot :-)).
Usually, a high scoring move with low samples is noise. But you might as well invest just a little more effort and see if the high scoring move remains high scoring. If it's just noise and not really a great move, it will die very quickly. So even a modest time extension makes sense. In tactical situation where the new move really has been discovered to be best, the samples will very quickly ramp up. I think it's also perfectly feasible to choose the best scoring move that is sampled more than N percent of the highest sampled move. You use this if you don't want to mess around with the time control algorithm. It would be interesting for you to try this too. I would try something between 30 percent and 80 percent. I don't have a sense of which value might work best. Perhaps 50% is a good try. My guess is that it's better to extend the time, because it is a way to be smarter about time allocation - spending extra effort where it has more chance to pay off. - Don _______________________________________________ computer-go mailing list [EMAIL PROTECTED] http://www.computer-go.org/mailman/listinfo/computer-go/
