On Thu, 13 Jan 2011, Olivier Teytaud wrote:
2) discard simulations
- spot semeais
- solve them with a dedicated solver
- discard simulations which propose a result which is not
consistent with the dedicated solver
As when I've tested that I was almost sure it would work, I'd like to
explain why it does not :-)
Let's consider that a white group A is clearly alive. The semeai is
detected, the solver sees that
The white group A should not die. What happens ?
When we discard simulations which kill white group A (which should be
alive), then ok, simulations look better. However, they're better, but
they introduce a bias: white will play possible 10 moves for protecting
group A, whereas 3 stones are enough.
Understanding this took me a lot of time, but it's a clear explanation of
the result...
This conditioning idea is very interesting.
And the explanation for it failing just as much.
It shows that for the winning side, the right sequence is really a
maximum, and all deviations are a loss, while it is not for the losing
side.
In fact, this applies to all life-and-death...
Now, at first I would have feared that this system just discard too many
simulations. So I would have thought it's logical to combine it with
approach three, as you have done in the paper.
Didn't that also ensure that not too many moves are played at one place?
Something trigger-like really seems in order. With an interdiction to
play in the local area if the opposite side does not play there on the
previous move, or at the extreme end of the playout...
By the way, there was an article suggesting to use "meta"-nodes in the
tree (my vocabulary), that is nodes that would contain many possible
moves, and then the next node would be the real move chosen.
It might be interesting to use that with one node for each local area
(moves may belong to several areas), a priori with strict areas (that is
there would often be only one local area, the whole board, and local
would only be very clear-cut such as walled semeai) ? Moreover this
could allow to collect statistics for the local fights from the tree.
Jonas
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
