This post is to propose a metric that measures the effectiveness of a
playout policy
in a MC tree search. It could give some idea as how the playing strength
varies with
the total playout number.

Let N be the total playout number. The effectve search depth is defined as

Depth = (log with base f) (N/m),
where m is related to factors such as the threshold value used, etc. f is
the more
interestng number characterizing a playout policy. A playout that selects
moves
randomly gives the largest value of f. I thnk it could be 2 or bigger. For
the most
effective search policy available today, such as those used by the most
strong Go programs
at present is about 1.5.

So what can above calculation tell us? According to above calculation it
could estimate
that the effective search depth of the today's strong Go programs are about
11, if the playout
number is one million and assume m=600, f=1.5. If an effective search depth
of 50
is requied to reach high dan level. Then the playout number needs to
increase by a
factor of 1.5^39, about 7.4 milliom times. That is 7.4 trillion playouts s
neeeded.

DL
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to