Hi Petr,
Thanks for the great comments, sorry to be so slow in getting back to
you (on vacation/workshop...)
Hello,
On Sun, Apr 06, 2008 at 08:55:26PM -0600, David Silver wrote:
> Here is a draft of the paper, any feedback would be very welcome :-)
>
> http://www.cs.ualberta.ca/~silver/research/publications/files/MoGoNectar.pdf
you are saying that in minimax, opponent moves are selected by
minimizing the lower confidence bound - this seems novel, is that
so? I
always got the impression that for the opponent moves, you reverse the
mean value but still use UCB.
I think this just depends on how you formulate the problem. If you use
minimax then the opponent moves should minimise a lower confidence
bound. If you use negamax then the opponent moves should maximise an
upper confidence bound.
I see three unclear details about the RAVE algorithm: Are only nodes
in the tree considered for the inclusion, or also moves in the
following
random playout?
Also moves in the following random playout. Will clarify, thanks.
And one of the sentences hints that there is a separate
period of playouts purely seeding the RAVE value before the UCT-RAVE
linear combination takes over - is that so?
No, there is no separate period, it is always a linear combination
(with decaying weight).
And it does not follow from
the paper that that the UCB formula is used for the RAVE value as
well,
while the ICML paper states that.
Unfortunately we did not have space for this discussion :-(
In any case, later versions of MoGo used (still use?) an exploration
constant of zero, which means that the UCB formula is no longer used.
I am surprised on the bad effect of the grandfather heuristic and
the
good effect of the even game heuristic. I assume that the effect of
the
heuristics should accumulate when several of them are combined in the
prior value?
Yes, I agree it is surprising! Combining sources of prior knowledge
may well be a good idea, I don't believe much work has been done on
this topic.
The paper looks very nice otherwise.
Thanks!
-Dave
PS Apologies for the pdf problems that some people encountered, I'll
post a new version on my website shortly._______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
