On Tue, 2007-05-01 at 08:35 -0400, Álvaro Begué wrote:
> We also chose (b), but not for the reason Chris posted. In chess you
> typically don't do any initial search to guess what the opponent will
> do; you use the second move from the PV line in the previous search.
I don't know what you do, but my UCT program generates a principal
variation, I always have a move to try without an extra search and
I usually have a PV several moves deep.
I guess every program is different. I chose method A because the
subtree that you are interested in using method B is always smaller
than if you predict the move correctly and just search directly.
In this case it's a pretty big win - you get 100% benefit from
the prediction.
If method A doesn't predict the move correctly, you get NO benefit
from method A. With method B you get a small subtree that can be
used, but when I looked at the numbers it was almost a useless
subtree. Of course I am talking about the case where method B
prefers a different subtree than what the opponent played.
When I considered the numbers it was pretty clear that method A
was better. I didn't play a hunch, I actually generated data
concerning move prediction rates and subtree sizes and analyzed
the data.
But like most everything else, you should check this for yourself if you
are interested in getting the most performance possible. It could vary
on other implementations.
- Don
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
