On Tue, Oct 27, 2009 at 08:47:41AM +0100, Olivier Teytaud wrote:
> > Could you give us at least a general picture of improvements compared to
> > what was last published as
> > www.lri.fr/~teytaud/eg.pdf<http://www.lri.fr/%7Eteytaud/eg.pdf>? Is it
> > "just"
> > further tuning and small tweaks or are you trying out some exciting new
> > things? ;-)
> >
>
> There is one important improvement, for which I must check with coauthors if
> they agree for me to explain it. Below the other recent improvements in 9x9.
>
> We have also recently encoded some (very simple) tricks against bad cases as
> we had
> against Fan Hui (i.e. cases in which the only good move is not simulated).
> Roughly,
> is the value of the node is very bad, then simulate randomly among the sons.
> We can
> show (mathematically) that with such tricks, we have the consistency (as
> UCT),
> plus some frugality (i.e. we do not simulate all the tree, even with
> infinite computation time whereas UCT simulates all the tree AND simulates
> all the tree infinitely often).
> It gives very little improvement in self-play, but it understands better at
> least the situation seen in the game with Fan Hui. What I like in this
> improvement is that it's
> the first time there is something which was mathematically developped for
> mogo and
> which leads to a positive result. Well, maybe this changes only 1% of games,
> but maybe it makes mogo more robust for complicated ko fights which do not
> occur in self-play.
Interesting! Conceptually, I don't like this that much since it just
work-arounds RAVE bias instead of solving it in more general way, but I
can see its technical value.
AIUI, once upon N simulations in a node you take let's say the node with
the lowest value, pick one son of it at random within the tree and start
a simulation?
> Finally, there was a GP-based development of new patterns. However, this is
> quite minor I guess - I like the fact that this GP-based module works in a
> somehow stable manner, but maybe it would only be worth using it on an
> implementation which is not
> yet optimized.
Wow - one of my planned little projects was genetic development of the
3x3 patterns... To evaluate patterns, do you use tournaments or some
smarter method? I feared one generation would take awfully long...
--
Petr "Pasky" Baudis
A lot of people have my books on their bookshelves.
That's the problem, they need to read them. -- Don Knuth
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
