There are two distinct concepts here. There is precision and there is accuracy. Your test will test precision (how exact the estimate is), but it fails to test accuracy (a measure of how close the mean is to the actual value). But, I do think that knowing the precision is useful. If the precision is too high (the noise is too low), it indicated that too many simulations (on a given position) were done for efficient use of processing power. Atleast that is my perception of the situation.
In a more general sense, I think that data gathering is definately by far one of the most useful ways to contribute to the computer go community and I salute you for your proposal. - Nick On 6/16/07, Heikki Levanto <[EMAIL PROTECTED]> wrote:
It seems generally accepted that MC or UCT programs are weakest in the opening. My own experience matches this too. Some times I get the idea that my program doesn't know at all what it is doing the first few moves. I propose a simple test to see if that is the case. Before doing it, I'd like to hear if other programmers are willing to make a similar test, and agree on the particulars. The idea is simple: We know that the empty board is symmetrical. How many iterations does it take before the program gets values for the first move that are symmetrical too. For example, if the program prefers the 3-3 point in one corner, but 4-4 in the other corner, then I would say that it doesn't know much about the situation. I am sure there is a mathematically sound way to measure how symmetric the evaluation is, but my math is a bit rusty, so I am asking if someone can come up with a good way. After that, I'm asking if various programmers would be willing to run this test, and publish the results? - Heikki (who has no time to debug his own program...) -- Heikki Levanto "In Murphy We Turst" heikki (at) lsd (dot) dk _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
