On 9/24/07, Don Dailey <[EMAIL PROTECTED]> wrote: > > > I've been toying around with time control lately (and failing to get it > > right). > > The main principle in time control is that early moves are far more > important than later moves. When 2 equal fairly weak bots play, the > game is usually over (for all practical purposes) in the first very few > moves. Often in the first 10 moves for example. So you might as well > spend a lot more time on those moves.
A very simple and yet effective strategy is to simply allocate a > fraction of the remaining time (as reported by the GTP time_left > command) to the next move. The fraction you use should be based on the > board size. This will cause early moves to be allocated much more > effort than later moves and ALSO has the desirable quality of speeding > up the end-games. > > Arrange it so that the first move (that is not a book move) is allocated > considerably more time than the 30th or 40th move. > > I put a lower bound on how low the allocated time can dip - because when > it get's to the point where it is playing really fast, the game is very > likely to be within a few moves of completion. But the lower bound is > still pretty low - it's less than the 1/4 second CGOS gift and it's only > purpose is to prevent a nearly random move. > > You can try to get more sophisticated, but you will probably be doing a > lot of work for a gain you can't measure. I would be surprised if you > could find a better time control that could be demonstrated to be better > and it certainly won't be simpler. > > KISS is almost always best - if not KISS you better have a really good > reason. > I had a very kiss algorithm previously. It was very close to time_to_use = 0.1 * (time_left - 60) I noticed that most of the time, the 60 second buffer wasn't getting used. I've been experimenting with a reduced buffer size as a function of the number of moves... time_to_use = 0.1 * (time_left - x*(120-move_num) -1)+0.1, where 0.1<x<1. The results from this have been relatively promising, but I'm trying too hard to get x small. I've been trying to make x adaptive based on self-timing (which is where I really deviate from KISS).
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
