>My general impression (also based on experiences from chess): >Distributing time rather balanced over the moves is a stable >strategy.
Reasoning on the basis of experience in chess is OK, but you must account for the differences between the domains. Chess is more or less uniformly difficult across the whole game. Go is not. It is definitely more difficult in the opening, especially for MCTS programs. Trials take longer in the opening, and the variance is larger, and the differences between moves is smaller (usually) which means that fewer moves are obviously forced. You have to spend more time on early moves in MCTS Go programs. Pebbles calculates the time required to uniformly spread the remaining time over the game. It then *doubles* that amount, and allocates that much time for the current play. This policy is not as extreme as you might think; it results in more-or-less uniform numbers of trials across the whole game. I have some experimental evidence that suggests that doubling is not enough. Perhaps the optimum multiplier is 2.5 or 3. Now, this usually does not result in having to play blitz moves later in the game. (It can happen, if the opponent drags out a losing effort into 100+ turns, but that doesn't matter.) Mogo might have gone too far, but maybe not. There are a lot of ways to lose games. Best, Brian _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
