Hi Christoph,
Thanks for replying.
Christoph Birk wrote:
On Feb 15, 2008, at 3:29 AM, Tim Foden wrote:
In your "pure MC program", do you use UCB1 to choose the next move to
search at the root? If not, what algorithm are you using? I'm
currently using UCB1 for my test in Fluke.
No, it uses a random move even at the root node.
myCtest does NOT create a tree (except for the UCT version). That
probably
explains why myCtest gains so much (more than Fluke) for addng AMAF
Well there was a bug in the version of Fluke that I ran initially
(080214)... I wasn't checking if the AMAF moves were valid in the root
node. Once I fixed this it gained another 120 elo. Now it's rated
about 65 elo higher than myCtest-10k-AMAF. I guess I'd expect this now
that I know that you're choosing moves randomly at the root, but Fluke
is using UCB1 to select moves at the root.
'myCtest-xxk-AMAF-y' has only one addition. It assigns the random-
playout result not only to the first move, but to a fraction of all
moves (simple all-moves-as-first). > What's different about the -3,
-5, and -8 versions? There seems to be > significant strength
differences between them. using N := moves in random simulation a :=
AMAF moves a = pow(N,y) with y = 0.25 (3) y = 0.50 (5) y = 0.75 (8)
due to the name-length limitation of CGOS (18 characters) Christoph
A question about the AMAF version... When you are updating the
statistics for the AMAF moves... which moves are you using? e.g. are
you using just the moves of the side to move (e.g. move 1, move 3,
move 5.... move n), or every move (e.g. move 1, move 2, move 3...
move n), or some other thing?
I use of course only the moves of 'this' player , I guess that
corresponds to
move 1,3,5,...n in you example.
Yes. I only tried using all of them as I vaguely remember a comment
somewhere about a good move tending to be good for both sides. But in
my experiment this was obviously worse. It's nice to get the
confirmation though. :)
When I tried it with every move Fluke seemed to play noticeably
weaker than with every other move.
I would expect that. If you use moves that were not played by the
winning/losing side that should mix up the statistics of good/bad moves
and create a somehow random move.
I also notice that there's a myCtest-10k-AMAF, with no '-y' on the
end. Is this the same as the one with '-8', or is there some other
change?
myCtest-AMAF is the final version ... I think it uses the 0.75 exponent.
Thanks. It seems to me that it may be worth trying 0.7 and 0.8 to see
if 0.75 is a maximum or not.
Cheers, Tim.
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
