On 9, Jul 2007, at 4:01 PM, Sylvain Gelly wrote:
Hi,
> I'm on the other side of this issue. In my opinion all kinds of go
> knowledge are fair game and I'm rather disappointed that so small
> amounts of domain specific knowledge have been merged with the UCT
> search approaches.
>
I agree. I really do not understand why using domain knowledge would
be a problem for some people.
Maybe one big reason is that domain knowledge is really hard to get
right. A lot of man power is needed to have a program with non
ridiculous go knowledge, higher than any other ideas that have been
successful in UCT.
I agree with you that it is hard. It is easy to include simple domain
knowledge that is only sometimes right and often wrong. It is hard
start with one set of domain knowledge and add to it in a way that
improves play.
It seems to me that this opinion is held
by programmers with less Go knowledge who hope that pure search is
an answer.
Maybe they believe that there is a lot of room for improvement without
go knowledge, and other will put the go knowledge because they are
stronger in that part. Isn't it the way research goes forward?
Absolutely. I just think that the UCT method has given us the biggest
part of what it can without biasing methods that stay away from random
games and select moves that are closer to the relevant space. This is
easier in thermal physics, where the theory tells us that Boltzman
distributions are what we need. You are closer to these methods and
I am just starting to learn them, so you may be more correct than I.
(...) I cannot imagine that progress will be
made without a great deal of domain knowledge.
Depending on what you exactly mean I disagree.
I mean progress by the standard usually applied to computer Go:
programs that can beat 1D humans on a full board, and then get
better.
Progress has been made
without "a great deal of domain knowledge" and there are many
improvements in the algorithms we can make. That does not prevent
using domain knowledge!
I think that the
next big trick will be in getting the domain knowledge into the form
that the MC methods can use efficiently.
I agree that it is ONE OF the next big trick. But people with
experience in the "classical" programs are much more suited for that
than "new comers". On the other side, there are plenty of other ways
to improve.
My point is that they are not mutually exclusive.
Cheers,
Sylvain
We agree that work on algorithms as well as on data representation
are needed. I am less sure that knowledge representation in the
classical
programs is the right expertise for its representation in the MC
playouts.
So many thing yet to do!
Cheers,
David
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
