Hi!
On Fri, Mar 08, 2013 at 12:30:22AM +0000, Aja Huang wrote:
> > Now it seems to me that this is related to the way playouts are done
> > and it will be difficult to improve with Mogo style (rule-based)
> > playouts above certain strength, without using larger patterns and next
> > move choice based on probability distribution. Currently, playing out
> > a simple joseki in a sensible way in simulations will just never happen.
> > This is a bit frustrating since all my attempts at successfully
> > implementing probdist-based playouts have failed so far, but I guess
> > I will just have to try again...
>
>
> To implement softmax, you can refer to my thesis where I have described the
> framework of the move generator for the playout. Detecting forbidden moves
> and replacing useless moves by better alternatives are very useful. There
> you can gain a lot by applying much Go-knowledge. Two good candidate
> algorithms for training the feature weights are MM and SB(Simulation
> Balancing). I tried hard but failed to measure any improvement from SB
> gammas (trained on 9x9) on 19x19. You can use CLOP to tune the MM gammas
> which are far from optimal according to our experience.
Thank you for the reference. It's true that in my experiments, I don't
follow the "forbid - replace" logic but rather apply this logic when
assigning the features; your idea is nice as it should be significantly
more efficient, though I will have to rework my code quite a bit in
order to accomodate it.
A question that has always been important for me is how wide set of
features to match and how often to recompute them. I assume that you are
mainly matching local tactical features and 3x3 patterns. When there is
no local move, do you choose a global move randomly or do you constuct
a probability distribution for the whole board?
Thanks,
--
Petr "Pasky" Baudis
For every complex problem there is an answer that is clear,
simple, and wrong. -- H. L. Mencken
_______________________________________________
Computer-go mailing list
Computer-go@dvandva.org
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
