Álvaro Begué wrote:
There are many things in the paper that we had never thought of, like
considering the distance to the penultimate move.
That feature improved the effectiveness of progressive widening a lot.
When I had only the distance to the previous move, and the opponent made
a dangerous move, Crazy Stone sometimes wanted to tenuki at the root, to
move the focus of local search away from the danger. Considering the
distance to the penultimate move fixes a lot of the problems of local
search. Maybe I should try to consider distances to even older moves.
Also the approach you devised with John seems to be very similar to what
I do, indeed. I also worked on the idea of adjusting the random policy
online to adapt to the current position, with a method that looks like
incremental Elo rating algorithms. I am deeply convinced that this can
bring huge improvements. If the random policy only uses general
patterns, whatever tree search we do at the root won't be able to make a
strong Go player. We have to find a way to influence random simulations,
not only inside the tree near the root, but also far from the root. Once
we know how to do this effectively, we'll have very strong programs.
Rémi
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
