Hi Peter, On 5/2/07, Peter Drake <[EMAIL PROTECTED]> wrote:
Orego also uses option B. Because UCT eventually focuses search on the most promising moves, it probably will spend most of its time on a single move, effectively doing A without the need for extra parameter settings.
Yes, this is one of the reasons why I thought B might be better. One of the main benefits in using option A is that you easily save a lot of time for your remaining moves when you get a ponder hit. Using method B, do you adjust the amount of time to think by the size of the subtree generated by the ponder search? ie, if you want to search for 10 seconds, and your program performs 10K samples a second, then you expect about ~100K samples for the move. If you already start with 95K (because of the pondering), do you adjust your thinking time to roughly perform an extra 5K samples? Joel _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
