Isaac Deutsch wrote:
Hi,
The "criticality" stuff looks really interesting. Do you apply it with the
offline knowledge, or as a RAVE prior value, or otherwise? It looks like
you precalculate (before the MTCS) the ownership + criticality map, maybe
it can be extracted from playouts in the MTCS as well?
ibd
I use it as a pattern feature in my pattern-learning scheme described in
the Amsterdam 2007 paper.
I am not sure what you mean by "offline knowledge" or "RAVE prior value".
My system computes a "gamma" value for each move. This value is used in
two ways:
- progressive widening: moves with bad gamma are pruned
- progressive bias: the value of a move, instead of w/n is (w+c*gamma)
/ n, with c a constant.
I don't precalculate before MCTS. Criticality, like point owner, is
re-calculated for each node of the tree, as it grows. It is computed
from MCTS playouts.
More precisely, search at a node starts without MC features. Move gammas
are all computed with static features only. After a few playouts, gammas
are updated with MC features, computed from the outcome of the first MC
playouts.
Rémi
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
