[Computer-go] effectiveness of transposition tables for go
gschmidt958 at yahoo.com
Tue May 11 09:59:19 PDT 2010
Thomas Lavergne wrote:
>So when I select a child I use only the winrate of choosing this specific child from this specific position, not choosing any child from any goban who have lead to the same target position. I think this is cleaner.
Is this what you're referring to? In Kocsis's paper, he warns against the following:
"It would be tempting to use the number of simulations
that lead to the evaluation of the transposition for the bias
term as well. However, this could lead to the value of
the parent converging to incorrect value, if most of the
simulations for the position that follows the best move in
the position is reached frequently through different paths (i.e.
when N s,a /Ns is much lower than N g(s,a)/Ns)."
So in your implementation, I assume the parent node maintain the actual counts through this node for this parent alone? In other words, you use N s,a as opposed to the larger N g(s,a)?
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Computer-go